CN111028253B - Method and device for dividing fine iron powder - Google Patents
Method and device for dividing fine iron powder Download PDFInfo
- Publication number
- CN111028253B CN111028253B CN201911164824.1A CN201911164824A CN111028253B CN 111028253 B CN111028253 B CN 111028253B CN 201911164824 A CN201911164824 A CN 201911164824A CN 111028253 B CN111028253 B CN 111028253B
- Authority
- CN
- China
- Prior art keywords
- convolution
- attention
- module
- result
- carrying
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/12—Edge-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10024—Color image
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30108—Industrial image inspection
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Life Sciences & Earth Sciences (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a method and a device for segmenting fine iron powder, which can completely and accurately extract fine iron powder from fine iron powder images. The method comprises the following steps: acquiring an iron fine image and a corresponding label graph in the transmission process as a training sample set; establishing a segmentation network SegNet based on an attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention; training a segmentation network SegNet by using the acquired training sample set; and segmenting the iron ore concentrate image to be segmented by using a trained segmentation network SegNet, and outputting the category of each pixel point of the iron ore concentrate image. The invention relates to the technical field of ore screening.
Description
Technical Field
The invention relates to the technical field of ore screening, in particular to a method and a device for dividing iron fine powder.
Background
The iron concentrate is a raw material for producing pellets, and the iron content determines the quality of the iron concentrate. In China, mineral powder is rich in resources in China, has large reserves, has a plurality of small particles and complex varieties, and causes that the mineral powder resources in China are rich in yield but unsatisfactory in quality, and the balling property and the stability of components are affected. Therefore, most of mineral powder adopted by various domestic enterprises is imported, but the quantity of imported mineral is increased along with the expansion of demands. Due to limited iron ore resources, the supply of raw materials is insufficient and has to be turned to the domestic market. Meanwhile, the steel industry of China is developed towards high quality and finish machining, and the quality of the iron concentrate serving as a raw material must be suitable for the variety.
The traditional technology for analyzing the taste of the fine iron powder generally uses a chemical or physical method, and the time and labor costs are relatively high. Deep learning is a machine learning method which is widely and novel in application, is a high-efficiency algorithm, utilizes features extracted from different levels as a recognition basis to obtain the category, can effectively solve the defects of time consumption, labor consumption and the like of the traditional fine iron powder, and can well guarantee that higher accuracy is achieved at lower cost and maintenance cost is reduced.
Because of the similarity of the shape, color, texture and volume of the iron concentrate and other ores mixed therein, some of the iron concentrate is difficult to distinguish by naked eyes; in addition, some iron concentrates and ores are very small in volume, and the area presented in the picture occupies very few pixels, which is a small target. In addition, the small target has less information, so that the information contained in the small target is easy to ignore in the pooling and convolution processes, and meanwhile, the existing semantic segmentation method is relatively difficult to extract edge characteristics (due to inaccurate edge positioning) and small target information, so that the classification accuracy of the edge pixel points and the pixel points with small areas is poor in the classification accuracy of the pixel points in a large range, and the complete extraction of the target object of the fine iron powder is not facilitated.
Disclosure of Invention
The invention aims to solve the technical problems of small characteristic difference, too small target volume or inaccurate edge positioning between iron concentrate and other ores mixed in the iron concentrate in the prior art, so that the iron concentrate cannot be completely and accurately extracted.
In order to solve the above technical problems, an embodiment of the present invention provides a method for cutting fine iron powder, including:
acquiring an iron fine image and a corresponding label graph in the transmission process as a training sample set;
establishing a segmentation network SegNet based on an attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention;
training a segmentation network SegNet by using the acquired training sample set;
and segmenting the iron ore concentrate image to be segmented by using a trained segmentation network SegNet, and outputting the category of each pixel point of the iron ore concentrate image.
Further, before training the segmentation network SegNet using the acquired training sample set, the method further comprises:
carrying out consistent segmentation processing on the obtained iron oxide fine image and the corresponding label graph;
and carrying out image enhancement processing on the iron ore concentrate image after the segmentation processing and the corresponding label graph.
Further, the image enhancement processing includes: one or more of rotation, flipping, zooming in, zooming out, and translation.
Further, the split network SegNet includes: an encoder and a decoder based on an attention mechanism module;
the encoder is used for carrying out convolution, batch normalization and pooling processing on the input image and extracting a feature map;
the decoder is used for extracting the spatial attention characteristic from the characteristic diagram output by the encoder, upsampling, extracting the channel attention characteristic, performing splicing, convolution and batch normalization processing on the extracted spatial attention characteristic and the channel attention characteristic, and outputting the category of each pixel point of the fine iron powder image.
Further, the decoder includes: the system comprises a first attention mechanism module, a first convolution normalization module connected with the first attention mechanism module, a second attention mechanism module connected with the first convolution normalization module, a second convolution normalization module connected with the second attention mechanism module, a third attention mechanism module connected with the second convolution normalization module, a third convolution normalization module connected with the third attention mechanism module, a fourth attention mechanism module connected with the third convolution normalization module, a fourth convolution normalization module connected with the fourth attention mechanism module, and a classifier module connected with the fourth convolution normalization module;
wherein each attention mechanism module comprises: the device comprises a space attention unit, an up-sampling layer connected with the space attention unit, a channel attention unit connected with the up-sampling layer and a splicing unit connected with the channel attention unit;
each convolution normalization module comprises: a plurality of network structures connected by convolutional layers and bulk normalization layers, wherein each convolutional layer is followed by a linear rectification function ReLU.
Further, the spatial attention unit is configured to perform convolution and deconvolution operations on the received feature graphs output by the adjacent batch normalization layers, perform convolution operations on the deconvolution result and the feature graphs output by the batch normalization layers corresponding to the deconvolution result and having the same number of features in the encoder, add the feature graphs, perform linear rectification functions ReLU, convolution, sigmoid and upsampling operations on the addition result, multiply the upsampling result with the feature graphs output by the batch normalization layers corresponding to the addition result and having the same number of features in the encoder, and perform convolution and batch normalization operations on the multiplication result to extract spatial attention features, where sigmoid represents a hyperbolic tangent S-type function.
Further, the channel attention unit is configured to perform convolution operation on an output result of the up-sampling layer, perform compression operation on a convolution result by using the global pooling layer, model a correlation between channels by using two full-connection layers according to the compression result, recover feature dimensions by adjusting dimensions and shapes of the matrix, multiply a feature map of the recovered dimensions with the convolution result, and extract channel attention features.
Further, the training the segmentation network SegNet by using the acquired training sample set includes:
and training the segmentation network SegNet by using the obtained training sample set, and optimizing the segmentation network SegNet by taking a loss function of the minimized segmentation network SegNet as an objective function.
Further, the loss function of the split network SegNet is:
wherein L is fl Representing a loss value; alpha represents a balance factor; gamma represents a regulatory factor; y' represents a predicted output value; y represents the label of the training sample.
The embodiment of the invention also provides a device for cutting the fine iron powder, which comprises the following components:
the acquisition module is used for acquiring the iron powder images and the corresponding annotation pictures in the transmission process as a training sample set;
the establishing module is used for establishing the segmentation network SegNet based on the attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention;
the training module is used for training the segmentation network SegNet by using the acquired training sample set;
the segmentation module is used for segmenting the iron ore concentrate image to be segmented by utilizing the trained segmentation network SegNet and outputting the category of each pixel point of the iron ore concentrate image.
The technical scheme of the invention has the following beneficial effects:
in the scheme, in order to acquire the characteristics contained in the small target and more characteristic information which is different from other ores, channel attention is increased on the basis of the original segmentation network SegNet, and in order to acquire more edge information and better positioning target objects, space attention is increased on the basis of the original segmentation network SegNet, so that the trained segmentation network SegNet based on the space attention and the channel attention is utilized to completely and accurately extract the iron concentrate from the iron concentrate images to be segmented in the acquired transmission process, thereby improving the extraction precision and the working efficiency of the iron concentrate and achieving the purpose of generating high-quality pellet raw materials.
Drawings
Fig. 1 is a schematic flow chart of a method for dividing fine iron powder according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of a SegNet of a split network based on an attention mechanism module according to an embodiment of the present invention;
FIG. 3 is a schematic diagram of a workflow of a spatial attention unit according to an embodiment of the present invention;
FIG. 4 is a schematic diagram of a workflow of a channel attention unit according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an apparatus for separating fine iron powder according to an embodiment of the present invention.
Detailed Description
In order to make the technical problems, technical solutions and advantages to be solved more apparent, the following detailed description will be given with reference to the accompanying drawings and specific embodiments.
Aiming at the problems that the existing iron concentrate powder and other ores mixed in the iron concentrate powder have small characteristic difference, small target volume or inaccurate edge positioning, the iron concentrate powder cannot be completely and accurately extracted, the invention provides an iron concentrate powder segmentation method and a segmentation device.
Example 1
As shown in fig. 1, the method for cutting fine iron powder provided by the embodiment of the invention comprises the following steps:
s101, acquiring an iron powder image and a corresponding annotation graph in the transmission process as a training sample set;
s102, establishing a segmentation network SegNet based on an attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention;
s103, training a segmentation network SegNet by using the acquired training sample set;
s104, segmenting the iron ore concentrate image to be segmented by using a trained segmentation network SegNet, and outputting the category of each pixel point of the iron ore concentrate image.
According to the method for segmenting the iron concentrate, in order to obtain the characteristics contained in the small targets and more characteristic information which is different from other ores, channel attention is increased on the basis of the original segmentation network SegNet, and in order to obtain more edge information and better positioning target objects, space attention is increased on the basis of the original segmentation network SegNet, so that the trained segmentation network SegNet based on the space attention and the channel attention is utilized to completely and accurately extract the iron concentrate from the iron concentrate images to be segmented in the acquired transmission process, the extraction precision and the working efficiency of the iron concentrate are improved, and the purpose of generating high-quality pellet raw materials is achieved.
In this embodiment, segNet is a deep convolutional encoder-decoder architecture for image segmentation.
In a specific embodiment of the foregoing method for segmentation of fine iron powder, further, before training the segmentation network SegNet using the obtained training sample set, the method further includes:
carrying out consistent segmentation processing on the obtained iron oxide fine image and the corresponding label graph;
and carrying out image enhancement processing on the iron ore concentrate image after the segmentation processing and the corresponding label graph.
In the embodiment, firstly, an iron powder image in the transmission process can be obtained through a high-definition industrial camera, then, a part of iron powder images are selected, manual labeling is carried out through a field expert, a labeling chart of the selected iron powder images is obtained, and the selected iron powder images and the corresponding labeling chart are used as a training sample set; and taking other iron ore concentrate images as a test sample set, wherein the test sample set is used for testing the accuracy of segNet segmentation of the trained segmentation network.
In this embodiment, it is assumed that the size of the obtained original fine iron powder image is 2592×1944, the size of the image after image segmentation is set to 256×256, and the corresponding label map of each training sample image is also subjected to consistent segmentation processing.
In this embodiment, the image enhancement processing includes: one or more of rotation, flipping, zooming in, zooming out, translating; for example, randomly rotating the image by a certain angle, changing the orientation of the image content; flipping the image in a horizontal or vertical direction; enlarging or reducing the image according to a certain proportion; the image is translated over the image plane.
In a specific embodiment of the foregoing method for splitting fine iron powder, the splitting network SegNet further includes: an encoder (encocoder) and an attention mechanism module-based decoder (encocoder);
the encoder is used for carrying out convolution, batch normalization and pooling processing on the input image and extracting a feature map;
the decoder is used for extracting the spatial attention characteristic from the characteristic diagram output by the encoder, upsampling, extracting the channel attention characteristic, performing splicing, convolution and batch normalization processing on the extracted spatial attention characteristic and the channel attention characteristic, and outputting the category of each pixel point of the fine iron powder image.
In this embodiment, the feature is extracted in an encoder, and the decoder is used to analyze the extracted feature.
In this embodiment, as shown in fig. 2, a brief description will be given of an encoder:
1) The encoder includes: 5 convolution normalization modules (1, 2, 3, 4, 5) and 4 maximum pooling layers max pool (1, 2, 3, 4), each convolution normalization module comprising: a plurality of network structures connected by convolutional layers (conv) and bulk normalization layers (BN), each convolutional layer being followed by a linear rectification function (Rectified Linear Unit, reLU), wherein ReLU is an activation function;
2) The convolution normalization modules 1 and 2 respectively comprise 2 network structures connected by convolution layers and batch normalization layers; the convolution normalization modules 3, 4 and 5 respectively comprise 3 network structures connected by convolution layers and batch normalization layers; thus, the encoder has a total of 13 convolutional layers;
3) The convolution normalization module 1 is connected with the maximum pooling layer 1, the maximum pooling layer 1 is connected with the convolution normalization module 2, the convolution normalization module 2 is connected with the maximum pooling layer 2, the maximum pooling layer 2 is connected with the convolution normalization module 3, the convolution normalization module 3 is connected with the maximum pooling layer 3, the maximum pooling layer 3 is connected with the convolution normalization module 4, the convolution normalization module 4 is connected with the maximum pooling layer 4, and the maximum pooling layer 4 is connected with the convolution normalization module 5;
4) The convolution layer is used for extracting features of the image, the batch normalization layer is used for relieving the problem of gradient dispersion in a deep network, and the pooling layer can gradually reduce network training parameters and avoid segNet over fitting of a segmentation network. The number of feature channels is gradually increased and the size of the feature map is gradually reduced due to convolution and pooling operations in the encoder, so as to extract more accurate features.
In this embodiment, as shown in fig. 2, a decoder will be briefly described:
1) The decoder includes: the system comprises a first attention mechanism module, a first convolution normalization module connected with the first attention mechanism module, a second attention mechanism module connected with the first convolution normalization module, a second convolution normalization module connected with the second attention mechanism module, a third attention mechanism module connected with the second convolution normalization module, a third convolution normalization module connected with the third attention mechanism module, a fourth attention mechanism module connected with the third convolution normalization module, a fourth convolution normalization module connected with the fourth attention mechanism module, and a classifier module (conv+softmax) connected with the fourth convolution normalization module; that is, the decoder includes 4 attention mechanism modules (first, second, third, fourth), 4 convolution normalization modules (first, second, third, fourth), and 1 classifier module, wherein each convolution normalization module includes: a plurality of network structures connected by convolution layers and batch normalization layers, each convolution layer being followed by a linear rectification function ReLU;
2) Each of the attention mechanism modules includes: a spatial Attention (Spatial Attention) unit, an upsampling layer (upsampling) connected to the spatial Attention unit, a Channel-wise Attention (Channel-wise Attention) unit connected to the upsampling layer, a concatenation unit (concatenation) connected to the Channel-wise Attention unit, so that the decoder uses 4 upsampling layers;
in this embodiment, spatial Attention may increase extraction of spatial features, and Channel-wise Attention may increase extraction of Channel features. Adding Spatial Attention units in front of each upsampling layer, adding Channel-wise Attention units after each upsampling layer, splicing the features extracted by the space Attention units and the Channel Attention units by a splicing unit, and transmitting the features to the later convolution layers; thus, the network can obtain more context information, and the separation result is more accurate. It should be noted that the size and the number of channels of the two feature maps to be spliced are required to be consistent;
3) The first, second and third convolution normalization modules respectively comprise 3 network structures connected by convolution layers and batch normalization layers; the fourth convolution normalization module comprises 2 network structures connected by convolution layers and batch normalization layers; thus, the encoder has a total of 11 convolutional layers;
4) The up-sampling layer up-samples the set feature size layer by layer, in the decoder, after each up-sampling, the number of feature channels is halved, and the size of the feature map is doubled, so that the original resolution size image can be restored.
5) The classifier module consists of conv and softmax and is used for processing the output of the fourth convolution normalization module to generate nonlinear transformation and outputting the category of each pixel point of the iron ore concentrate image.
In a specific embodiment of the foregoing method for partitioning fine iron powder, as shown in fig. 3, the spatial attention unit is further configured to perform convolution and deconvolution (conv transform) operation on the received feature map (y) output by the adjacent batch normalization layer, perform convolution operation on the deconvolution result and the feature map (z) output by the batch normalization layer with the same number of features corresponding to the deconvolution result in the encoder, perform addition (add), extract more features from the addition result by using a linear rectification function ReLU, convolution, hyperbolic tangent S-type function (sigmoid) and upsampling operation, and then Multiply (multiplex) the upsampling result and the feature map (z) output by the batch normalization layer with the same number of features corresponding to the upsampling result in the encoder, that is, add a weight to each feature, so that important information of a local picture can be extracted by transformation; and then, performing convolution and batch normalization operation on the multiplication result, and extracting the spatial attention characteristic.
In the foregoing specific embodiment of the method for partitioning fine iron powder, as shown in fig. 4, the channel attention unit is further configured to perform convolution operation on the output result (y) of the upsampling layer, perform compression operation on the convolution result by using the global pooling layer, model the correlation between channels according to the compression result by using the full-connection layer (dense) 1 and the full-connection layer 2, restore the feature dimension by adjusting the dimension and shape (Reshape) of the matrix (i.e., firstly, reduce the feature dimension to 1/16 of the input through the full-connection layer 1, then, after activated by ReLU, return to the original dimension through the full-connection layer 2 and Reshape operation), output and input the weights with the same number of features, multiply the feature map and the convolution result of the restored dimension, weight to the feature of each channel, and extract the channel attention feature. By doing so, the extracted channel attention features have more nonlinearities, so that complex correlations among channels are better fitted, and the number of parameters and the calculation amount are greatly reduced.
In a specific embodiment of the foregoing method for segmentation of fine iron powder, further, training the segmentation network SegNet using the obtained training sample set includes:
and training the segmentation network SegNet by using the obtained training sample set, and optimizing the segmentation network SegNet by taking a loss function of the minimized segmentation network SegNet as an objective function.
In this embodiment, the loss function of the segment network SegNet is:
wherein y represents the mark of the fine iron powder.
Wherein L is fl Representing a loss value; alpha represents a balance factor for balancing the number proportion non-uniformity of the positive and negative samples; gamma represents the adjustment factor, making the separation network more focused on difficult, misclassified samples; y 'represents a predicted output value of the segment network SegNet, and the value of y' is between 0 and 1; y denotes the label of the training sample, 1 denotes the ferrite powder sample, and 0 denotes the background sample.
In this embodiment, the positive sample refers to a sample belonging to a certain category, the negative sample refers to a sample not belonging to a certain category, and the positive sample is the ferrite fine powder sample, and the negative sample is the background sample.
In this embodiment, alpha and gamma are both adjustable super parameters, and y' is a componentThe prediction result of the cut network SegNet has a value between (0-1), when y=1, y' tends to be 1, representing a positive sample which is easy to identify, and its contribution to the weight tends to be 0; when y=0, y' tends to 0, representing an easily identifiable negative sample, its contribution to the weight tends to 0. Thus, the loss function L fl The problem of unbalanced categories can be solved, and not only the weight of the background category is reduced, but also the weight of easily identified positive/negative samples (easy positive/negative) is reduced.
In this embodiment, the segmented network SegNet may be optimized using an optimizer random gradient descent (stochastic gradient descent, SGD), where the learning rate is 0.1, the learning rate decay value (decay) is set to 1e-4, the momentum parameter (momentum) is 0.9, and momentum indicates how much to preserve the original update direction, which is between 0-1.
In this embodiment, the to-be-segmented iron powder image is segmented and subjected to image enhancement, and then the trained model is used to segment the to-be-segmented iron powder image after image enhancement, and the class of each pixel point is output, wherein the class is represented by 0 and 1, 0 represents the background, and 1 represents the iron powder.
Example two
The present invention also provides a specific embodiment of the apparatus for dividing fine iron powder, and since the apparatus for dividing fine iron powder provided by the present invention corresponds to the specific embodiment of the method for dividing fine iron powder, the apparatus for dividing fine iron powder can achieve the object of the present invention by executing the steps of the flow in the specific embodiment of the method, and therefore the explanation in the specific embodiment of the method for dividing fine iron powder is also applicable to the specific embodiment of the apparatus for dividing fine iron powder provided by the present invention, and will not be repeated in the following specific embodiments of the present invention.
As shown in fig. 5, an embodiment of the present invention further provides an apparatus for dividing fine iron powder, including:
the acquisition module 11 is used for acquiring the iron powder image and the corresponding annotation graph in the transmission process as a training sample set;
an establishing module 12, configured to establish a SegNet based on an attention mechanism module, where the attention mechanism includes: spatial attention and channel attention;
a training module 13, configured to train the segmentation network SegNet by using the obtained training sample set;
the segmentation module 14 is configured to segment the iron ore concentrate image to be segmented by using a trained segmentation network SegNet, and output a class of each pixel point of the iron ore concentrate image.
In order to acquire the characteristics contained in a small target and more characteristic information different from other ores, the device for segmenting the iron concentrate increases channel attention on the basis of the original segmentation network SegNet, and increases spatial attention on the basis of the original segmentation network SegNet in order to acquire more edge information and better locate target objects, so that the trained segmentation network SegNet based on the spatial attention and the channel attention is utilized to completely and accurately extract the iron concentrate from the iron concentrate images to be segmented in the acquired transmission process, thereby improving the extraction precision and the working efficiency of the iron concentrate and achieving the purpose of generating high-quality pellet raw materials.
It is noted that relational terms such as first and second, and the like are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions.
While the foregoing is directed to the preferred embodiments of the present invention, it will be appreciated by those skilled in the art that various modifications and adaptations can be made without departing from the principles of the present invention, and such modifications and adaptations are intended to be comprehended within the scope of the present invention.
Claims (6)
1. The method for cutting the fine iron powder is characterized by comprising the following steps of:
acquiring an iron fine image and a corresponding label graph in the transmission process as a training sample set;
establishing a segmentation network SegNet based on an attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention;
training a segmentation network SegNet by using the acquired training sample set;
dividing the iron ore concentrate image to be divided by using a trained segmentation network SegNet, and outputting the category of each pixel point of the iron ore concentrate image;
wherein the split network SegNet comprises: an encoder and a decoder based on an attention mechanism module;
the encoder is used for carrying out convolution, batch normalization and pooling processing on the input image and extracting a feature map;
the decoder is used for extracting spatial attention features, upsampling, extracting channel attention features, splicing, convoluting and normalizing the extracted spatial attention features and channel attention features on the feature map output by the encoder in batches, and outputting the category of each pixel point of the fine iron powder image;
wherein the decoder comprises: the system comprises a first attention mechanism module, a first convolution normalization module connected with the first attention mechanism module, a second attention mechanism module connected with the first convolution normalization module, a second convolution normalization module connected with the second attention mechanism module, a third attention mechanism module connected with the second convolution normalization module, a third convolution normalization module connected with the third attention mechanism module, a fourth attention mechanism module connected with the third convolution normalization module, a fourth convolution normalization module connected with the fourth attention mechanism module, and a classifier module connected with the fourth convolution normalization module;
wherein each attention mechanism module comprises: the device comprises a space attention unit, an up-sampling layer connected with the space attention unit, a channel attention unit connected with the up-sampling layer and a splicing unit connected with the channel attention unit;
each convolution normalization module comprises: a plurality of network structures connected by convolution layers and batch normalization layers, wherein each convolution layer is followed by a linear rectification function ReLU;
the spatial attention unit is used for carrying out convolution and deconvolution operation on the received characteristic graphs output by the adjacent batch normalization layers, carrying out convolution operation on the deconvolution result and the characteristic graphs output by the batch normalization layers with the same number of characteristics corresponding to the deconvolution result in the encoder, adding the deconvolution result, carrying out linear rectification function ReLU, convolution, sigmoid and up-sampling operation on the added result, carrying out multiplication on the up-sampling result and the characteristic graphs output by the batch normalization layers with the same number of characteristics corresponding to the added result in the encoder, carrying out convolution and batch normalization operation on the multiplication result, and extracting the spatial attention characteristics, wherein sigmoid represents a hyperbolic tangent S-type function;
the channel attention unit is used for carrying out convolution operation on the output result of the up-sampling layer, carrying out compression operation on the convolution result by utilizing the global pooling layer, modeling the correlation between channels by utilizing two full-connection layers according to the compression result, recovering the characteristic dimension by utilizing the operation of adjusting the dimension and the shape of the matrix, carrying out multiplication on the characteristic diagram of the recovered dimension and the convolution result, and extracting the channel attention characteristic.
2. The method of claim 1, further comprising, prior to training the segmentation network SegNet with the acquired set of training samples:
carrying out consistent segmentation processing on the obtained iron oxide fine image and the corresponding label graph;
and carrying out image enhancement processing on the iron ore concentrate image after the segmentation processing and the corresponding label graph.
3. The iron nugget segmentation method of claim 2, wherein the image enhancement process comprises: one or more of rotation, flipping, zooming in, zooming out, and translation.
4. The method of claim 1, wherein training the segmentation network SegNet using the obtained training sample set comprises:
and training the segmentation network SegNet by using the obtained training sample set, and optimizing the segmentation network SegNet by taking a loss function of the minimized segmentation network SegNet as an objective function.
5. The method of claim 4, wherein the loss function of the SegNet is:
wherein L is fl Representing a loss value; alpha represents a balance factor; gamma represents a regulatory factor; y, representing a predicted output value; y represents the label of the training sample.
6. An iron concentrate segmenting device, characterized by comprising:
the acquisition module is used for acquiring the iron powder images and the corresponding annotation pictures in the transmission process as a training sample set;
the establishing module is used for establishing the segmentation network SegNet based on the attention mechanism module, wherein the attention mechanism comprises: spatial attention and channel attention;
the training module is used for training the segmentation network SegNet by using the acquired training sample set;
the segmentation module is used for segmenting the iron ore concentrate image to be segmented by utilizing a trained segmentation network SegNet and outputting the category of each pixel point of the iron ore concentrate image;
wherein the split network SegNet comprises: an encoder and a decoder based on an attention mechanism module;
the encoder is used for carrying out convolution, batch normalization and pooling processing on the input image and extracting a feature map;
the decoder is used for extracting spatial attention features, upsampling, extracting channel attention features, splicing, convoluting and normalizing the extracted spatial attention features and channel attention features on the feature map output by the encoder in batches, and outputting the category of each pixel point of the fine iron powder image;
wherein the decoder comprises: the system comprises a first attention mechanism module, a first convolution normalization module connected with the first attention mechanism module, a second attention mechanism module connected with the first convolution normalization module, a second convolution normalization module connected with the second attention mechanism module, a third attention mechanism module connected with the second convolution normalization module, a third convolution normalization module connected with the third attention mechanism module, a fourth attention mechanism module connected with the third convolution normalization module, a fourth convolution normalization module connected with the fourth attention mechanism module, and a classifier module connected with the fourth convolution normalization module;
wherein each attention mechanism module comprises: the device comprises a space attention unit, an up-sampling layer connected with the space attention unit, a channel attention unit connected with the up-sampling layer and a splicing unit connected with the channel attention unit;
each convolution normalization module comprises: a plurality of network structures connected by convolution layers and batch normalization layers, wherein each convolution layer is followed by a linear rectification function ReLU;
the spatial attention unit is used for carrying out convolution and deconvolution operation on the received characteristic graphs output by the adjacent batch normalization layers, carrying out convolution operation on the deconvolution result and the characteristic graphs output by the batch normalization layers with the same number of characteristics corresponding to the deconvolution result in the encoder, adding the deconvolution result, carrying out linear rectification function ReLU, convolution, sigmoid and up-sampling operation on the added result, carrying out multiplication on the up-sampling result and the characteristic graphs output by the batch normalization layers with the same number of characteristics corresponding to the added result in the encoder, carrying out convolution and batch normalization operation on the multiplication result, and extracting the spatial attention characteristics, wherein sigmoid represents a hyperbolic tangent S-type function;
the channel attention unit is used for carrying out convolution operation on the output result of the up-sampling layer, carrying out compression operation on the convolution result by utilizing the global pooling layer, modeling the correlation between channels by utilizing two full-connection layers according to the compression result, recovering the characteristic dimension by utilizing the operation of adjusting the dimension and the shape of the matrix, carrying out multiplication on the characteristic diagram of the recovered dimension and the convolution result, and extracting the channel attention characteristic.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911164824.1A CN111028253B (en) | 2019-11-25 | 2019-11-25 | Method and device for dividing fine iron powder |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201911164824.1A CN111028253B (en) | 2019-11-25 | 2019-11-25 | Method and device for dividing fine iron powder |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111028253A CN111028253A (en) | 2020-04-17 |
| CN111028253B true CN111028253B (en) | 2023-05-30 |
Family
ID=70206469
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201911164824.1A Active CN111028253B (en) | 2019-11-25 | 2019-11-25 | Method and device for dividing fine iron powder |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111028253B (en) |
Families Citing this family (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112950586A (en) * | 2021-03-02 | 2021-06-11 | 攀钢集团攀枝花钢铁研究院有限公司 | LF furnace steel slag infrared identification method and system |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109325911A (en) * | 2018-08-27 | 2019-02-12 | 北京航空航天大学 | A kind of space base rail detection method based on attention enhancing mechanism |
| CN110059717A (en) * | 2019-03-13 | 2019-07-26 | 山东大学 | Convolutional neural networks automatic division method and system for breast molybdenum target data set |
| CN110070073A (en) * | 2019-05-07 | 2019-07-30 | 国家广播电视总局广播电视科学研究院 | Pedestrian's recognition methods again of global characteristics and local feature based on attention mechanism |
| WO2019153908A1 (en) * | 2018-02-11 | 2019-08-15 | 北京达佳互联信息技术有限公司 | Image recognition method and system based on attention model |
| CN110197182A (en) * | 2019-06-11 | 2019-09-03 | 中国电子科技集团公司第五十四研究所 | Remote sensing image semantic segmentation method based on contextual information and attention mechanism |
-
2019
- 2019-11-25 CN CN201911164824.1A patent/CN111028253B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2019153908A1 (en) * | 2018-02-11 | 2019-08-15 | 北京达佳互联信息技术有限公司 | Image recognition method and system based on attention model |
| CN109325911A (en) * | 2018-08-27 | 2019-02-12 | 北京航空航天大学 | A kind of space base rail detection method based on attention enhancing mechanism |
| CN110059717A (en) * | 2019-03-13 | 2019-07-26 | 山东大学 | Convolutional neural networks automatic division method and system for breast molybdenum target data set |
| CN110070073A (en) * | 2019-05-07 | 2019-07-30 | 国家广播电视总局广播电视科学研究院 | Pedestrian's recognition methods again of global characteristics and local feature based on attention mechanism |
| CN110197182A (en) * | 2019-06-11 | 2019-09-03 | 中国电子科技集团公司第五十四研究所 | Remote sensing image semantic segmentation method based on contextual information and attention mechanism |
Non-Patent Citations (1)
| Title |
|---|
| 赵欣欣 ; 钱胜胜 ; 刘晓光 ; .基于卷积神经网络的铁路桥梁高强螺栓缺失图像识别方法.中国铁道科学.(第04期),全文. * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111028253A (en) | 2020-04-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111062915B (en) | Real-time steel pipe defect detection method based on improved YOLOv3 model | |
| CN111340123A (en) | Image score label prediction method based on deep convolutional neural network | |
| CN110827312B (en) | Learning method based on cooperative visual attention neural network | |
| CN111680705B (en) | MB-SSD method and MB-SSD feature extraction network suitable for target detection | |
| CN112819748B (en) | Training method and device for strip steel surface defect recognition model | |
| CN110991359A (en) | Satellite image target detection method based on multi-scale depth convolution neural network | |
| CN111127454A (en) | Method and system for generating industrial defect sample based on deep learning | |
| CN112819837B (en) | Semantic segmentation method based on multi-source heterogeneous remote sensing image | |
| CN113837193A (en) | Zinc flotation froth image segmentation algorithm based on improved U-Net network | |
| CN114612664A (en) | Cell nucleus segmentation method based on bilateral segmentation network | |
| CN112926486A (en) | Improved RFBnet target detection algorithm for ship small target | |
| CN111028253B (en) | Method and device for dividing fine iron powder | |
| CN115035371A (en) | Borehole wall crack identification method based on multi-scale feature fusion neural network | |
| CN115546466A (en) | Weak supervision image target positioning method based on multi-scale significant feature fusion | |
| CN116385374A (en) | Cell counting method based on convolutional neural network | |
| CN114092467A (en) | Scratch detection method and system based on lightweight convolutional neural network | |
| CN112784894B (en) | Automatic labeling method for rock slice microscopic image | |
| CN110175548A (en) | Remote sensing images building extracting method based on attention mechanism and channel information | |
| CN104050673B (en) | Depth map extraction method based on K-means image segmentation | |
| CN113436198A (en) | Remote sensing image semantic segmentation method for collaborative image super-resolution reconstruction | |
| CN112784806A (en) | Lithium-containing pegmatite vein extraction method based on full convolution neural network | |
| CN114494703B (en) | Intelligent workshop scene target lightweight semantic segmentation method | |
| CN116740572A (en) | Marine vessel target detection method and system based on improved YOLOX | |
| CN116091862A (en) | Picture quality identification method, device, equipment, storage medium and product | |
| CN113012167A (en) | Combined segmentation method for cell nucleus and cytoplasm |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |