CN113449622A - Image classification, identification and detection method for cotton plants and weeds - Google Patents
Image classification, identification and detection method for cotton plants and weeds Download PDFInfo
- Publication number
- CN113449622A CN113449622A CN202110682324.8A CN202110682324A CN113449622A CN 113449622 A CN113449622 A CN 113449622A CN 202110682324 A CN202110682324 A CN 202110682324A CN 113449622 A CN113449622 A CN 113449622A
- Authority
- CN
- China
- Prior art keywords
- noise
- image
- neural network
- layer
- convolutional neural
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- 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
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Processing (AREA)
- Image Analysis (AREA)
Abstract
The invention belongs to the field of image processing and computer vision, and discloses a cotton plant and weed image classification, identification and detection method, which is based on an anti-noise convolutional neural network and comprises the following steps: acquiring an image to be classified; a noise mapping layer that detects various types of noise and generates a noise map for each image to indicate noisy and uncorrupted images; the size of the layer is adaptively adjusted according to the noise point so as to improve the robustness of the network to the noise; noise map based data enhancement to generate additional data for the convolutional neural network; identifying and extracting image region characteristics by using a noise self-adaptive convolution layer; the noise self-adaptive pooling layer is used for down-sampling the characteristic diagram, reducing data dimensionality and improving learning speed; a full connection layer for realizing final classification and prediction; and outputting an image classification result. The invention can accurately identify and classify cotton plants and weeds and prepare for the next mechanical topping of cotton plants.
Description
Technical Field
The invention belongs to the field of image processing and machine vision, and particularly relates to an image classification, identification and detection method for cotton plants and weeds.
Background
Cotton is an important component of economic crops in China and is also a main raw material of textile industry in China. Topping is required during the cotton growth process, i.e., cutting the top core of the cotton when the cotton has grown to a certain height or period to increase the cotton yield. At present, cotton topping is mainly finished manually, and the cotton topping method is high in labor intensity and low in efficiency. Therefore, under the conditions of current labor shortage and high labor cost, the method has important practical significance and wide development prospect in quickly, accurately and efficiently realizing automatic topping of cotton, and therefore, the high-speed accurate classification and identification of cotton leaves and weeds is the premise and the basis of automatic topping.
The machine vision system converts the object to be shot into image signal through machine vision product (image shooting device, which is divided into CMOS and CCD), and transmits it to special image processing system, to obtain the form information of the object to be shot, and converts it into digital signal according to the information of pixel distribution, brightness and color, the image system carries out various operations to these signals to extract the characteristic of the object, and then controls the action of the equipment on site according to the result of the discrimination.
Image classification is used in a variety of applications, such as security, educational and promotional systems, and in recent years, much research has been conducted to design automated systems that extract essential features from images. The convolutional neural network is an effective image classification method, which uses convolution, pooling and full-link layers for learning, and is a multi-layer neural network composed of neurons with trainable weights and biases. The noise image is a destructive factor in the convolutional neural network training, the classification performance of the network is reduced, and the removal of noise from the image is an important problem in image processing, and generally as a preprocessing step, various types of noise, such as impulse noise, image sample loss, data packet loss in image transmission, image damage and tampering, affect the image quality, so that the image is not suitable for image processing, the impulse noise is one of the most common noise types in image acquisition, recording and transmission, the intensity of a pixel damaged by the impulse noise is higher or lower than that of an undamaged adjacent pixel, and when some parts of the image are lost, damaged or partially blocked by an unwanted object, the image sample loss occurs. Digital images can be tampered maliciously, the image quality is reduced, and the recovery of the tampered images is an expensive process, and the performance of a convolutional neural network is affected by the images polluted by noise. Noisy images are often recovered in a pre-processing stage, resulting in an improvement in the classification performance of the convolutional neural network, however, in addition, the preprocessing of noise removal is an expensive and time-consuming process, and a noise-robust convolutional neural network is proposed for classification of the noise image, which does not require denoising preprocessing, in the proposed method, for various types of noise such as impulse noise, lost image samples, packet loss in image transmission, corrupted images and tampered images, by adding a noise mapping layer and a self-adaptive adjustment size layer, the structure of the basic convolutional neural network is improved, the basic convolutional neural network is made to be robust to noise, and classification of noise images in different components is considered.
Disclosure of Invention
Aiming at the problem of accurate identification of cotton plants, the invention utilizes a machine vision technology and carries out classification and accurate identification on images of the cotton plants and weeds based on an anti-noise convolutional neural network, thereby providing preparation for subsequent cotton topping mechanical research.
In order to achieve the purpose, the technical scheme adopted by the invention is as follows.
An image classification, identification and detection method for cotton plants and weeds comprises the following steps:
and 7, outputting an image classification result.
Specifically, in step 2: a noise mapping layer is placed at the beginning of the anti-noise convolutional neural network and is used for detecting impulse noise, lost image samples, data packet loss in image transmission, damaged images and tampered images; the noise mapping layer detects various types of noise and generates a noise map for each image to indicate noisy and uncorrupted images; impulse noise is detected by a local consistency index method based on image content detection packet loss in image transmission; further, the target area is used to detect missing image samples, the statistical method is used to detect damaged images, and the image authentication method is used to detect tampered images; in the proposed method, in order to make the convolutional neural network robust to noise, a noise map is generated for each image based on the type of noise detected in the previous step, each image containing four channels: the method comprises a noise map channel, a red channel, a blue channel and a green channel, wherein images are given to a convolutional neural network by four channels, and noise pixels are processed by noise robust convolutional neural network learning in a noise map-based training process.
Specifically, in step 3: convolutional neural networks generally start with convolutional layers, the input image of which should generally have a fixed size (244 × 244), and therefore the size of the image larger than the size of the convolutional neural network input should be reduced; a new layer is added into the convolutional neural network, and the new layer is called a self-adaptive adjusting layer, so that the robustness of the network to noise is improved; the task of the adaptive adjustment layer is to improve the dimension reduction method of a larger image than the convolutional neural network input using a noise map, select pixels from several pixels for reducing the size of the image using the noise map, the noise pixels are removed, and perform size reduction using the remaining pixels; the adaptive adjustment layer makes the noise pixel not participate in the dimension reduction process of the input image of the convolutional neural network.
Data is added by changing the light intensity of the image based on the noise map:
where I is the input image, P is the noise map, X is the light intensity variation, and O is the output image. Given the values of the noise map (P), each pixel of the image takes into account the following conditions:
x relates to an integer in the interval [ -a, a ] that determines the brightness variation of each pixel, and the method changes the value of only the non-contaminated pixels, without considering the noisy pixels during the light intensity variation, which reduces the amount of calculation.
Specifically, in step 4: the convolution layer is a core part of the convolutional neural network, and the calculation and final integration are carried out on corresponding elements by covering an image area with the same size as that of the convolutional kernel, so that the main characteristic information of the image in the area is obtained, the spatial information of the original image area is completely reserved, and the position relation of adjacent pixels is not transformed or damaged; taking an image with the size of 5 × 5 as an example, selecting a convolution kernel of 3 × 3 to perform convolution operation on the image, wherein the process is actually performing sliding calculation on a matrix representing pixel values by a filter, and assuming that the distance of one pixel is moved each time, a calculation result of each sliding is the pixel value of a corresponding pixel point in an output target feature map.
The convolutional layer parameters comprise a group of filters which can be learned, the convolutional layer structure of the convolutional neural network is improved through self-adaptive filtering, and an effective method for improving the anti-noise capability of the convolutional neural network is provided; the method reduces the influence of noise on the convolutional neural network by removing the noise connection between the convolutional source and the convolutional kernel, and improves the classification precision by performing adaptive filtering with noise connection based on the pixel values in the neural network; discarding noisy connections prevents the noisy pixels from entering the next layer, and the proposed method of removing noisy connections can be used for different convolution kernel sizes, with the noise map updated at each layer;
the self-adaptive step size noise adjustment algorithm improves the classification precision of the noise image, and comprises the following steps: making a bit matrix as a step pitch graph; marking selected pixels in each stride; matching the noise map and the step map to detect noise pixels; the filter position is changed from the noise region to the nearest uncorrupted pixel in the w x w neighborhood.
Correcting the adaptive step map so that noise pixels are not considered in the classification process, initially matching the noise map and the step map together, and then performing a step operation to improve the classification of the noise image, w is the location area searched to find the nearest uncorrupted location for changing the filter's location.
Specifically, in step 5: the pooling layer has the main functions of down-sampling data, reducing the spatial dimension of the characteristic diagram, removing unimportant data information with low reference value, retaining most important information and further reducing the number of parameters; in a convolution system structure, a plurality of pooling layers can be placed between convolution layers, the width and height of an input image are reduced on the pooling layers so as to reduce the number of parameters and calculation, a pooling function replaces network output at a specific position with statistics and summary of adjacent output, noise directly influences the pooling layers, a new pooling operator correction method is provided for improving the accuracy of a convolution neural network in noise image classification, in order to eliminate the noise of the pooling layers, the pooling operator should avoid processing noise pixels, the selection of noise values in subsequent convolution neural network image layers is avoided, the accuracy of image classification is improved, and the proposed method of the noise pixel pool can be reused because the pooling operator is repeated for multiple times in the layers of the convolution neural network.
Specifically, in step 6: the full-connection layer can be placed behind the convolutional layer and the pooling layer, and the full-connection layer bears the main calculation amount of the convolutional neural network and is generally used as an output layer to output a result so as to realize final classification and prediction; all the neurons in full connection can be regarded as important characteristic information finally extracted by the network model, meanwhile, the adjacent network layers are in a full connection state, and each neuron or characteristic is measured by an independent weight, so that the network model has a comprehensive analysis function.
For the classification problem, the softmax function is usually applied to the output layer (the last layer of the convolutional neural network), the optimal values of the parameters of the convolutional neural network of the classification problem (e.g., weight vector and bias term) can be realized by minimizing the loss function, x (n) is expressed as the nth input data, y (n) is expressed as the real target label of the nth input data, o (n) is expressed as the nth output of the convolutional neural network classification, θ is expressed as all parameters, and the loss of the convolutional neural network can be obtained as:
the full-connection layer is a storage of final characteristic information of the convolutional neural network, and simultaneously plays a role in classification, all forms of characteristic expressions learned by the network are converted into an output space, and a final prediction result is obtained after data are integrated.
Due to the adoption of the technical scheme, compared with the prior art, the invention has the technical progress.
(1) The invention provides an anti-noise convolutional neural network, which is used for classifying noisy images without any denoising pretreatment, and improves the classification performance of the convolutional neural network on the noisy images;
(2) according to the invention, a noise mapping layer and a self-adaptive adjusting layer are added in the structure of the convolutional neural network, so that the convolutional layer, the convergence layer and the loss function of the convolutional neural network are improved, and the robustness of the neural network to noise is improved;
(3) the invention introduces the noise image-based adaptive data enhancement technology, improves the classification performance of a new algorithm, and improves the noise image classification and network training speed;
the invention belongs to the field of image processing and computer vision, and provides preparation for subsequent cotton topping machinery research by accurately identifying images of cotton plants and weeds based on an anti-noise convolutional neural network.
Drawings
FIG. 1 is a flow chart of a method of anti-noise convolutional neural network according to an embodiment of the present invention.
Fig. 2 is a structural diagram of a noise mapping layer according to an embodiment of the present invention.
Fig. 3 is a data enhancement architecture based on noise maps according to an embodiment of the present invention.
Fig. 4 shows the structure of convolutional layers proposed to enhance the robustness of the network to noisy images according to an embodiment of the present invention.
Fig. 5 is a workflow of a method for noise adaptation stride in convolutional neural networks of an embodiment of the present invention.
FIG. 6 is a 2 × 2 max pooling operator of an embodiment of the present invention and a step size of 2 pixels for undamaged pixels.
Fig. 7 shows the proposed improved method 2 × 2 max pooling operator with step size of 2 pixels to eliminate low density noise (black squares indicate noisy pixels) according to an embodiment of the present invention.
FIG. 8 shows a method for improving max-pooling operator to remove high density noise (black squares indicate noisy pixels) according to an embodiment of the present invention.
FIG. 9 is a flowchart illustrating the operation of the method for removing noise by the improved max-pooling operator according to an embodiment of the present invention.
Detailed Description
For the purpose of better explaining the present invention and to facilitate understanding, the present invention will be described in detail by way of specific embodiments with reference to the accompanying drawings.
The embodiment discloses an image classification, identification and detection method for cotton plants and weeds.
As shown in fig. 1, the present embodiment includes the following steps:
A noise mapping layer is placed at the beginning of the anti-noise convolutional neural network and is used for detecting impulse noise, lost image samples, data packet loss in image transmission, damaged images and tampered images; the noise mapping layer detects various types of noise and generates a noise map for each image to indicate noisy and uncorrupted images; the structure of the noise mapping layer is shown in fig. 2, and it can be seen from fig. 2 that impulse noise is detected by a local consistency index method based on image content to detect packet loss in image transmission; further, the target area is used to detect missing image samples, the statistical method is used to detect damaged images, and the image authentication method is used to detect tampered images; in the proposed method, in order to make the convolutional neural network robust to noise, a noise map is generated for each image based on the type of noise detected in the previous step, each image containing four channels: the method comprises a noise map channel, a red channel, a blue channel and a green channel, wherein images are given to a convolutional neural network by four channels, and noise pixels are processed by noise robust convolutional neural network learning in a noise map-based training process.
Convolutional neural networks generally start with convolutional layers, the input image of which should generally have a fixed size (244 × 244), and therefore the size of the image larger than the size of the convolutional neural network input should be reduced; a new layer is added into the convolutional neural network, and the new layer is called a self-adaptive adjusting layer, so that the robustness of the network to noise is improved; the task of the adaptive adjustment layer is to improve the dimension reduction method of a larger image than the convolutional neural network input using a noise map, select pixels from several pixels for reducing the size of the image using the noise map, the noise pixels are removed, and perform size reduction using the remaining pixels; the adaptive adjustment layer makes the noise pixel not participate in the dimension reduction process of the input image of the convolutional neural network.
Data is added by changing the light intensity of the image based on the noise map:
where I is the input image, P is the noise map, X is the light intensity variation, and O is the output image. Given the values of the noise map (P), each pixel of the image takes into account the following conditions:
x relates to an integer in the interval [ -a, a ] that determines the brightness variation of each pixel, and the method only changes the value of the non-contaminated pixels without considering the noise pixels during the light intensity variation, which reduces the amount of computation, and fig. 3 shows a data enhancement architecture based on a noise map.
And 4, identifying and extracting image region characteristics by the noise self-adaptive convolution layer.
The convolution layer is a core part of the convolutional neural network, and the calculation and final integration are carried out on corresponding elements by covering an image area with the same size as that of the convolutional kernel, so that the main characteristic information of the image in the area is obtained, the spatial information of the original image area is completely reserved, and the position relation of adjacent pixels is not transformed or damaged; taking an image with the size of 5 × 5 as an example, selecting a convolution kernel of 3 × 3 to perform convolution operation on the image, wherein the process is actually performing sliding calculation on a matrix representing pixel values by a filter, and assuming that the distance of one pixel is moved each time, a calculation result of each sliding is the pixel value of a corresponding pixel point in an output target feature map.
The convolutional layer parameters comprise a group of filters which can be learned, the convolutional layer structure of the convolutional neural network is improved through self-adaptive filtering, and an effective method for improving the anti-noise capability of the convolutional neural network is provided; the method reduces the influence of noise on the convolutional neural network by removing the noise connection between the convolutional source and the convolutional kernel, and improves the classification precision by performing adaptive filtering with noise connection based on the pixel values in the neural network; as shown in fig. 4, which is the proposed architecture of convolutional layers to enhance the robustness of the network to noisy images, discarding noisy connections prevents noisy pixels from entering the next layer, and the proposed method of removing noisy connections can be used for different convolutional kernel sizes, with the noise map updated at each layer.
The self-adaptive step size noise adjustment algorithm improves the classification precision of the noise image, and comprises the following steps: making a bit matrix as a step pitch graph; marking selected pixels in each stride; matching the noise map and the step map to detect noise pixels; the filter position is changed from the noise region to the nearest uncorrupted pixel in the w x w neighborhood.
Correcting the adaptive step map so that noise pixels are not considered in the classification process, initially matching the noise map and the step map together, and then completing the step operation, fig. 5 shows the workflow of the proposed method for noise adaptive step in convolutional neural networks to improve the classification of noise images, w × w is the location area searched to find the nearest uncorrupted location for changing the filter's location.
And 5, a noise self-adaptive pooling layer performs down-sampling on the feature map, reduces data dimensionality and improves the learning speed.
The pooling layer has the main functions of down-sampling data, reducing the spatial dimension of the characteristic diagram, removing unimportant data information with low reference value, retaining most important information and further reducing the number of parameters; in the convolution system structure, several pooling layers can be placed between convolution layers, the width and height of the input image are reduced by the pooling layers, so as to reduce the number of parameters and calculation, the pooling function replaces the network output of a specific position with the statistic summary of adjacent output, the noise directly affects the pooling layers, in order to improve the accuracy of the convolution neural network in the noise image classification, a new method for correcting the pooling operator is provided, a numerical value of the maximum pool operator is shown as figure 6, in order to eliminate the noise of the pooling layers, the pooling operator should avoid processing the noise pixel, figure 7 shows a proposed method for improving the maximum pooling operator to eliminate the low-density noise, the noisy pixel is shown by a black block, the pooling operator avoids processing the noise pixel, and avoids selecting the noise value for use in the subsequent convolution neural network layer, the accuracy of image classification is improved, and the proposed method of noise pixel pool can be repeatedly used because the pooling operator is repeatedly performed in the layer of the convolutional neural network; FIG. 8 shows a proposed method for improving the max-pooling operator to handle high-density noise pixels; fig. 9 shows the workflow of the proposed method for improving the maximal pooling operator for removing noise, in which one switch control is selected based on the pooling method of pixel values in the convolutional neural network.
And 6, fully connecting layers to realize final classification and prediction.
The full-connection layer can be placed behind the convolutional layer and the pooling layer, and the full-connection layer bears the main calculation amount of the convolutional neural network and is generally used as an output layer to output a result so as to realize final classification and prediction; all the neurons in full connection can be regarded as important characteristic information finally extracted by the network model, meanwhile, the adjacent network layers are in a full connection state, and each neuron or characteristic is measured by an independent weight, so that the network model has a comprehensive analysis function.
For the classification problem, the softmax function is usually applied to the output layer (the last layer of the convolutional neural network), the optimal values of the parameters of the convolutional neural network of the classification problem (e.g., weight vector and bias term) can be realized by minimizing the loss function, x (n) is expressed as the nth input data, y (n) is expressed as the real target label of the nth input data, o (n) is expressed as the nth output of the convolutional neural network classification, θ is expressed as all parameters, and the loss of the convolutional neural network can be obtained as:
the full-connection layer is a storage of final characteristic information of the convolutional neural network, and simultaneously plays a role in classification, all forms of characteristic expressions learned by the network are converted into an output space, and a final prediction result is obtained after data are integrated.
And 7, outputting an image classification result, accurately identifying and classifying the cotton plants and the weeds, and preparing for mechanical topping of the cotton plants in the next step.
Claims (6)
1. An image classification, identification and detection method for cotton plants and weeds is characterized by comprising the following steps:
step 1, obtaining an image to be classified;
step 2, a noise mapping layer detects various types of noise and generates a noise map for each image to indicate noisy and undamaged images;
step 3, the size of the layer is adjusted in a self-adaptive mode according to the noise point so as to improve the robustness of the network to the noise; noise map based data enhancement to generate additional data for the convolutional neural network;
step 4, identifying and extracting image region characteristics by using a noise self-adaptive convolution layer;
step 5, a noise self-adaptive pooling layer performs down-sampling on the feature map, reduces data dimensionality and improves learning speed;
step 6, fully connecting layers to realize final classification and prediction;
and 7, outputting an image classification result.
2. The image classification, identification and detection method for cotton plants and weeds as claimed in claim 1, wherein in step 2: a noise mapping layer is placed at the beginning of the anti-noise convolutional neural network and is used for detecting impulse noise, lost image samples, data packet loss in image transmission, damaged images and tampered images; the noise mapping layer detects various types of noise and generates a noise map for each image to indicate noisy and uncorrupted images; impulse noise is detected by a local consistency index method based on image content detection packet loss in image transmission; further, the target area is used to detect missing image samples, the statistical method is used to detect damaged images, and the image authentication method is used to detect tampered images; in the proposed method, in order to make the convolutional neural network robust to noise, a noise map is generated for each image based on the type of noise detected in the previous step, each image containing four channels: the method comprises a noise map channel, a red channel, a blue channel and a green channel, wherein images are given to a convolutional neural network by four channels, and noise pixels are processed by noise robust convolutional neural network learning in a noise map-based training process.
3. The image classification, identification and detection method for cotton plants and weeds as claimed in claim 1, wherein in step 3: convolutional neural networks generally start with convolutional layers, the input image of which should generally have a fixed size (244 × 244), and therefore the size of the image larger than the size of the convolutional neural network input should be reduced; a new layer is added into the convolutional neural network, and the new layer is called a self-adaptive adjusting layer, so that the robustness of the network to noise is improved; the task of the adaptive adjustment layer is to improve the dimension reduction method of a larger image than the convolutional neural network input using a noise map, select pixels from several pixels for reducing the size of the image using the noise map, the noise pixels are removed, and perform size reduction using the remaining pixels; the self-adaptive adjustment layer ensures that noise pixels do not participate in the dimension reduction process of the input image of the convolutional neural network; data is added by changing the light intensity of the image based on the noise map:
where I is the input image, P is the noise map, X is the light intensity variation, and O is the output image;
given the values of the noise map (P), each pixel of the image takes into account the following conditions:
x relates to an integer in the interval [ -a, a ] that determines the brightness variation of each pixel, and the method changes the value of only the non-contaminated pixels, without considering the noisy pixels during the light intensity variation, which reduces the amount of calculation.
4. The image classification, identification and detection method for cotton plants and weeds as claimed in claim 1, characterized in that in step 4: the convolution layer is a core part of the convolutional neural network, and the calculation and final integration are carried out on corresponding elements by covering an image area with the same size as that of the convolutional kernel, so that the main characteristic information of the image in the area is obtained, the spatial information of the original image area is completely reserved, and the position relation of adjacent pixels is not transformed or damaged; taking an image with the size of 5 × 5 as an example, selecting a convolution kernel of 3 × 3 to perform convolution operation on the image, wherein the process is actually performing sliding calculation on a matrix representing pixel values by a filter, and assuming that the distance of one pixel is moved every time, the calculation result of each sliding is the pixel value of a corresponding pixel point in an output target characteristic diagram;
the convolutional layer parameters comprise a group of filters which can be learned, the convolutional layer structure of the convolutional neural network is improved through self-adaptive filtering, and an effective method for improving the anti-noise capability of the convolutional neural network is provided; the method reduces the influence of noise on the convolutional neural network by removing the noise connection between the convolutional source and the convolutional kernel, and improves the classification precision by performing adaptive filtering with noise connection based on the pixel values in the neural network; discarding noisy connections prevents the noisy pixels from entering the next layer, and the proposed method of removing noisy connections can be used for different convolution kernel sizes, with the noise map updated at each layer;
the self-adaptive step size noise adjustment algorithm improves the classification precision of the noise image, and comprises the following steps: making a bit matrix as a step pitch graph; marking selected pixels in each stride; matching the noise map and the step map to detect noise pixels; changing the filter position from the noise region to the nearest uncorrupted pixel in the w × w neighborhood;
correcting the adaptive step map so that noise pixels are not considered in the classification process, initially matching the noise map and the step map together, and then performing a step operation to improve the classification of the noise image, w is the location area searched to find the nearest uncorrupted location for changing the filter's location.
5. The image classification, identification and detection method for cotton plants and weeds as claimed in claim 1, wherein in step 5: the pooling layer has the main functions of down-sampling data, reducing the spatial dimension of the characteristic diagram, removing unimportant data information with low reference value, retaining most important information and further reducing the number of parameters; in a convolution system structure, a plurality of pooling layers can be placed between convolution layers, the width and height of an input image are reduced on the pooling layers so as to reduce the number of parameters and calculation, a pooling function replaces network output at a specific position with statistics and summary of adjacent output, noise directly influences the pooling layers, a new pooling operator correction method is provided for improving the accuracy of a convolution neural network in noise image classification, in order to eliminate the noise of the pooling layers, the pooling operator should avoid processing noise pixels, the selection of noise values in subsequent convolution neural network image layers is avoided, the accuracy of image classification is improved, and the proposed method of the noise pixel pool can be reused because the pooling operator is repeated for multiple times in the layers of the convolution neural network.
6. The image classification, identification and detection method for cotton plants and weeds as claimed in claim 1, characterized in that in step 6: the full-connection layer can be placed behind the convolutional layer and the pooling layer, and the full-connection layer bears the main calculation amount of the convolutional neural network and is generally used as an output layer to output a result so as to realize final classification and prediction; all the neurons in full connection can be regarded as important characteristic information finally extracted by the network model, meanwhile, the adjacent network layers are in a full connection state, and each neuron or characteristic is measured by independent weight, so that the network model has a comprehensive analysis function;
for the classification problem, the softmax function is usually applied to the output layer (the last layer of the convolutional neural network), the optimal values of the parameters of the convolutional neural network of the classification problem (e.g., weight vector and bias term) can be realized by minimizing the loss function, x (n) is expressed as the nth input data, y (n) is expressed as the real target label of the nth input data, o (n) is expressed as the nth output of the convolutional neural network classification, θ is expressed as all parameters, and the loss of the convolutional neural network can be obtained as:
the full-connection layer is a storage of final characteristic information of the convolutional neural network, and simultaneously plays a role in classification, all forms of characteristic expressions learned by the network are converted into an output space, and a final prediction result is obtained after data are integrated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110682324.8A CN113449622A (en) | 2021-06-20 | 2021-06-20 | Image classification, identification and detection method for cotton plants and weeds |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110682324.8A CN113449622A (en) | 2021-06-20 | 2021-06-20 | Image classification, identification and detection method for cotton plants and weeds |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113449622A true CN113449622A (en) | 2021-09-28 |
Family
ID=77811902
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110682324.8A Pending CN113449622A (en) | 2021-06-20 | 2021-06-20 | Image classification, identification and detection method for cotton plants and weeds |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113449622A (en) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114359144A (en) * | 2021-12-01 | 2022-04-15 | 阿里巴巴（中国）有限公司 | Image detection method and method for obtaining image detection model |
| CN114612470A (en) * | 2022-05-10 | 2022-06-10 | 浙江浙能航天氢能技术有限公司 | Hydrogen-sensitive adhesive tape color change detection method based on improved image self-adaptive YOLO |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108520203A (en) * | 2018-03-15 | 2018-09-11 | 上海交通大学 | Multiple target feature extracting method based on fusion adaptive more external surrounding frames and cross pond feature |
| CN111915513A (en) * | 2020-07-10 | 2020-11-10 | 河海大学 | Image denoising method based on improved adaptive neural network |
| CN112580554A (en) * | 2020-12-25 | 2021-03-30 | 北京环境特性研究所 | CNN-based MSTAR data noise intensity control classification identification method |
-
2021
- 2021-06-20 CN CN202110682324.8A patent/CN113449622A/en active Pending
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108520203A (en) * | 2018-03-15 | 2018-09-11 | 上海交通大学 | Multiple target feature extracting method based on fusion adaptive more external surrounding frames and cross pond feature |
| CN111915513A (en) * | 2020-07-10 | 2020-11-10 | 河海大学 | Image denoising method based on improved adaptive neural network |
| CN112580554A (en) * | 2020-12-25 | 2021-03-30 | 北京环境特性研究所 | CNN-based MSTAR data noise intensity control classification identification method |
Non-Patent Citations (2)
| Title |
|---|
| MOHAMMAD MOMENY等: ""A noise robust convolutional neural network for image classification"", 《RESULTS IN ENGINEERING》, vol. 10, 9 June 2021 (2021-06-09), pages 1 - 12 * |
| 杨杰: "《人工智能基础》", 30 April 2020, 机械工业出版社, pages: 116 - 121 * |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114359144A (en) * | 2021-12-01 | 2022-04-15 | 阿里巴巴（中国）有限公司 | Image detection method and method for obtaining image detection model |
| CN114612470A (en) * | 2022-05-10 | 2022-06-10 | 浙江浙能航天氢能技术有限公司 | Hydrogen-sensitive adhesive tape color change detection method based on improved image self-adaptive YOLO |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108009542B (en) | Weed image segmentation method in rape field environment | |
| CN108562589A (en) | A method of magnetic circuit material surface defect is detected | |
| WO2022236876A1 (en) | Cellophane defect recognition method, system and apparatus, and storage medium | |
| CN113449622A (en) | Image classification, identification and detection method for cotton plants and weeds | |
| CN111178177A (en) | Cucumber disease identification method based on convolutional neural network | |
| CN113222959B (en) | Fresh jujube wormhole detection method based on hyperspectral image convolutional neural network | |
| CN113255434B (en) | Apple identification method integrating fruit characteristics and deep convolutional neural network | |
| CN105574514A (en) | Greenhouse immature tomato automatic identification method | |
| CN116630960B (en) | Corn disease identification method based on texture-color multi-scale residual shrinkage network | |
| CN111340019A (en) | Grain bin pest detection method based on Faster R-CNN | |
| CN117392627A (en) | Corn row line extraction and plant missing position detection method | |
| CN113723314A (en) | Sugarcane stem node identification method based on YOLOv3 algorithm | |
| CN110544237B (en) | Tea-oil tree plant disease and insect pest model training method and identification method based on image analysis | |
| CN111768456A (en) | Feature extraction method based on wood color space | |
| CN116110042A (en) | Tomato detection method based on CBAM attention mechanism of YOLOv7 | |
| CN113210264B (en) | Tobacco sundry removing method and device | |
| CN111667509B (en) | Automatic tracking method and system for moving target under condition that target and background colors are similar | |
| Yang et al. | Cherry recognition based on color channel transform | |
| Indukuri et al. | Paddy Disease Classifier using Deep learning Techniques | |
| Dhivya et al. | An Analysis Study of Various Image Preprocessing Filtering Techniques based on PSNR for Leaf Images | |
| Chen et al. | Rapid identification method of fresh tea leaves based on lightweight model | |
| CN117392440B (en) | Textile fabric retrieval method and system based on tissue structure and color classification | |
| CN114332442A (en) | Binocular color identification method integrated with learning mechanism | |
| CN113256671B (en) | Tree fruit counting method based on YCbCr color space | |
| CN117372881B (en) | Intelligent identification method, medium and system for tobacco plant diseases and insect pests |
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 | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20210928 |
|
| WD01 | Invention patent application deemed withdrawn after publication |