CN111833311B - Image recognition method based on deep learning and application of image recognition method in rice disease recognition - Google Patents
Image recognition method based on deep learning and application of image recognition method in rice disease recognition Download PDFInfo
- Publication number
- CN111833311B CN111833311B CN202010561990.1A CN202010561990A CN111833311B CN 111833311 B CN111833311 B CN 111833311B CN 202010561990 A CN202010561990 A CN 202010561990A CN 111833311 B CN111833311 B CN 111833311B
- Authority
- CN
- China
- Prior art keywords
- image
- deep learning
- training
- network
- trained
- 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
- 201000010099 disease Diseases 0.000 title claims abstract description 71
- 208000037265 diseases, disorders, signs and symptoms Diseases 0.000 title claims abstract description 71
- 238000000034 method Methods 0.000 title claims abstract description 65
- 238000013135 deep learning Methods 0.000 title claims abstract description 62
- 235000007164 Oryza sativa Nutrition 0.000 title claims abstract description 46
- 235000009566 rice Nutrition 0.000 title claims abstract description 46
- 240000007594 Oryza sativa Species 0.000 title abstract 2
- 238000012549 training Methods 0.000 claims abstract description 77
- 230000008569 process Effects 0.000 claims abstract description 18
- 238000012545 processing Methods 0.000 claims abstract description 9
- 230000003321 amplification Effects 0.000 claims abstract description 8
- 238000003199 nucleic acid amplification method Methods 0.000 claims abstract description 8
- 238000010276 construction Methods 0.000 claims abstract description 6
- 241000209094 Oryza Species 0.000 claims description 44
- 238000011176 pooling Methods 0.000 claims description 20
- 238000002372 labelling Methods 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 7
- 230000000007 visual effect Effects 0.000 claims description 7
- 230000009466 transformation Effects 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 5
- 238000007781 pre-processing Methods 0.000 claims description 4
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 claims description 3
- 238000005520 cutting process Methods 0.000 claims description 3
- RLLPVAHGXHCWKJ-UHFFFAOYSA-N permethrin Chemical compound CC1(C)C(C=C(Cl)Cl)C1C(=O)OCC1=CC=CC(OC=2C=CC=CC=2)=C1 RLLPVAHGXHCWKJ-UHFFFAOYSA-N 0.000 claims description 3
- 238000009827 uniform distribution Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 10
- 238000000605 extraction Methods 0.000 description 9
- 238000005457 optimization Methods 0.000 description 8
- 230000003902 lesion Effects 0.000 description 6
- 206010035148 Plague Diseases 0.000 description 5
- 241000607479 Yersinia pestis Species 0.000 description 5
- 241000208202 Linaceae Species 0.000 description 4
- 235000004431 Linum usitatissimum Nutrition 0.000 description 4
- 230000001580 bacterial effect Effects 0.000 description 4
- 230000000694 effects Effects 0.000 description 4
- 238000003709 image segmentation Methods 0.000 description 4
- 210000002569 neuron Anatomy 0.000 description 4
- 238000013528 artificial neural network Methods 0.000 description 3
- 235000013339 cereals Nutrition 0.000 description 3
- 238000005286 illumination Methods 0.000 description 3
- 238000003745 diagnosis Methods 0.000 description 2
- 238000010586 diagram Methods 0.000 description 2
- 238000013508 migration Methods 0.000 description 2
- 230000005012 migration Effects 0.000 description 2
- 238000010606 normalization Methods 0.000 description 2
- 230000011218 segmentation Effects 0.000 description 2
- 238000010187 selection method Methods 0.000 description 2
- 244000000231 Sesamum indicum Species 0.000 description 1
- 235000003434 Sesamum indicum Nutrition 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000013499 data model Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000002474 experimental method Methods 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 230000000644 propagated effect Effects 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 208000024891 symptom Diseases 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
-
- 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
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- 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
- G06T2207/10004—Still image; Photographic 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
- G06T2207/10—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
- G06T2207/20016—Hierarchical, coarse-to-fine, multiscale or multiresolution image processing; Pyramid transform
-
- 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]
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A40/00—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production
- Y02A40/10—Adaptation technologies in agriculture, forestry, livestock or agroalimentary production in agriculture
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Software Systems (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Mathematical Physics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses an image recognition method based on deep learning and application thereof in rice disease recognition, comprising the following steps: acquiring an image training set containing a target object; carrying out data enhancement processing on the training set image by adopting image amplification and image contrast adjustment; obtaining a trained deep learning network, wherein the trained deep learning network is obtained through training of an image training set and a constructed deep learning network to be trained, and the construction and training of the deep learning network to be trained are realized based on an auxiliary model; the method comprises the steps of obtaining an image to be identified, and identifying a target object in the image, wherein the method is completed by adopting an auxiliary model in the process of building and training a deep learning network, and the method is characterized in that the network model to be trained is built by selecting part of weight parameters and a network layer in the existing network model based on training of a big data set, and the training time is obviously shortened and the classification accuracy is improved by inputting an image training set to conduct network fine-tuning training.
Description
Technical Field
The invention relates to the field of image recognition, in particular to an image recognition method based on deep learning and application thereof to rice disease recognition.
Background
The diagnosis and identification of rice diseases are of great significance for improving rice quality. The image processing and machine vision technology is applied to the rice disease identification, has incomparable superiority compared with the traditional manual diagnosis and identification method, and improves the capability of crop disease monitoring and early warning.
In the rice disease identification process based on image identification, problems encountered include:
(1) The background information of the rice disease image is huge, and difficulty is brought to the segmentation process of a target area in the image;
(2) The rice diseases are various in types, and in the same growth period, the types and positions of the diseases are different. In different production periods, the shape, the color and other characteristics of the same kind of diseases are also different, the rice disease characteristics are changeable, the disease association degree is high, the disease complexity is high, the process of training and identifying by adopting a deep learning network model is adopted, the calculated amount and the modeling workload are increased, and the identifying speed is reduced;
(3) In the process of adopting the deep learning network model, a large number of rice disease images are required to be trained to achieve the identification accuracy, but the current rice disease images have no such huge number of data sets, and the ideal classification effect cannot be achieved.
Disclosure of Invention
Aiming at the problems existing in the prior art, the invention provides an image recognition method based on deep learning, which comprises the following steps:
(11) Acquiring an image training set containing a target object, wherein images in the training set carry annotation data;
(12) Carrying out data enhancement processing on the training set image by adopting image amplification and image contrast adjustment;
(13) Obtaining a trained deep learning network, wherein the trained deep learning network is obtained through training of an image training set and a constructed deep learning network to be trained, and the construction and training of the deep learning network to be trained are realized based on an auxiliary model;
(14) And acquiring an image to be identified, preprocessing, inputting the image into a training-completed deep learning network, and acquiring an identification result of a target object in the image.
As further optimization of the scheme, the auxiliary model selection method is that the task field of the auxiliary model is the same as the training task field of the image training set, and the similarity between the data set of the auxiliary model and the image training set meets a second preset threshold interval.
As a further optimization of the scheme, the similarity between the data set of the auxiliary model and the image training set is obtained by adopting a maximum mean difference algorithm (MMD).
As a further optimization of the above solution, the type of the auxiliary model employs a deep convolutional network.
As a further optimization of the above scheme, the method for constructing and training the deep learning network to be trained based on the auxiliary model is as follows:
(41) Acquiring a hidden layer and a hidden layer weight parameter of an auxiliary model;
(42) Migrating weight parameters of a convolution layer and a pooling layer of the auxiliary model and corresponding layers to a deep learning network to be trained;
(43) Adding a spatial pyramid pooling layer into a pooling layer in a deep learning network to be trained, fusing outputs of all the spatial pyramid pooling layers, and connecting to a full-connection layer, wherein the full-connection layer adopts 3 layers;
(44) The method comprises the steps of fixing weight parameters of a convolution layer and a pooling layer of a deep learning network to be trained, and inputting an image training set into the deep learning network to be trained to perform weight parameter training of a full-connection layer;
(45) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a first preset value, fixing the network parameters is canceled, and fine adjustment of the weight parameters of the whole network layer is performed;
(46) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a second preset value or the iteration number reaches a maximum preset value, stopping the training process.
As a further optimization of the scheme, the fully connected layers adopt 3 layers, and each fully connected layer adopts an L2 regularization and Dropout method.
As a further optimization of the scheme, the difference of the labeling data corresponding to the output value and the input value of the network model is measured by adopting a cross entropy function, and in the training process, the weight parameters of each layer are adjusted by carrying out counter propagation according to the cross entropy value at each output, and the adjustment method adopts an adam algorithm.
As a further optimization of the above scheme, the image amplification method includes:
changing the visual angle of the images in the image training set, firstly randomly selecting one of the images from the rotation angles of 0, 90, 180 and 270, and then randomly selecting and reversing the overturn of the input image to overturn along the x axis or along the y axis;
changing the size of the image, firstly cutting the image from the height and the width randomly by 0.1-0.4, and then carrying out random size scaling on the whole image by 0.2.
As a further optimization of the above scheme, the image contrast adjustment uses a uniform distribution for non-linearly converting the pixel value distribution of the original image to the first preset threshold interval, and the non-linear transformation formula:
wherein G is out For the pixel value of each pixel after transformation, G in G is the original pixel value exmin And G exmax To be the lower limit and the upper limit of the first preset threshold interval, G min And G max The average minimum pixel value and the maximum pixel value of the RGB color space are respectively calculated as follows:
G min ＝(G min ＝(R min +G min +B min )/3，G max ＝(R max +G max +B max )/3。
the invention also provides application of the image recognition method based on deep learning in rice disease recognition, wherein the images of the image training set comprise different kinds of disease images, different growth period images of the same kind of disease and the same kind of disease images of different rice areas.
The image recognition method based on deep learning and the application thereof in rice disease recognition have the following beneficial effects:
1. the image recognition method based on deep learning and the application thereof in rice disease recognition are suitable for recognition tasks with small image sample size and complex sample background, the construction and training process of a deep learning network are completed by adopting an auxiliary model, the network model which is completed by the existing training and is based on big data set training is utilized, part or all weight parameters and a network layer in the network model are selected to construct a network model to be trained, the training time is obviously shortened, the classification accuracy is improved by inputting an image training set, the complex background of a disease image of a rice disease image in a field is acquired, and the consideration is carried out, the image contrast adjustment is adopted to carry out data enhancement processing on the training set image, so that the target area and the background area in the image can be conveniently segmented in a deep neural network, and the recognition effect on four diseases of rice bacterial leaf blight, false smut, rice blast and flax leaf spot is realized.
2. According to the image recognition method based on deep learning and the application of the image recognition method in rice disease recognition, a layer of spatial pyramid pooling layer is added into the pooling layer in the deep learning network to be trained, outputs of all spatial pyramid pooling layers are fused and then connected to the full-connection layer, and the pooling layer adopts different sizes, so that the extracted image features have multi-scale, multi-scale feature extraction of rice diseases is realized, and disease recognition accuracy is improved.
Drawings
FIG. 1 is a block diagram of the overall flow of the image recognition method based on deep learning and application thereof in the recognition of rice diseases;
FIG. 2 is a block flow diagram of a method for constructing a deep learning network to be trained for application in the identification of rice diseases according to the image identification method based on deep learning of the invention;
fig. 3 is a flowchart of a training method of a deep learning network to be trained, which is based on the image recognition method of deep learning and the application of the method to the recognition of rice diseases.
Detailed Description
The technical scheme of the invention is further described below with reference to specific embodiments and drawings.
The invention provides an image recognition method based on deep learning, which is applied to rice disease recognition, and comprises the following steps:
(11) Acquiring an image training set containing a target object, wherein the images in the training set carry labeling data, the adopted images comprise different kinds of disease images, images of the same kind of disease in different growth periods and images of the same kind of disease in different rice areas, the diseases studied in the embodiment comprise four kinds of bacterial leaf blight, false smut, rice blast and flax leaf spot, and the following characteristics of the 4 kinds of disease spots are as follows:
(1) Bacterial leaf blight: the disease spots are usually arranged at the blade tip and the blade edge, then develop at the blade edge side or develop along the middle part of the blade, and are in yellow-white or off-white color and corrugated, and the two sides of the disease blade are bent towards the middle part;
(2) False smut: the disease part is ear part, which mainly affects grain, hypha in disease grain, after growing, hypha blocks become large, and then break, and yellow blocks visible to naked eyes are formed;
(3) Rice blast: the number of the hazard parts is large, the hazard period is long, and the hazard time and the hazard position are different and can be divided into 3 seedling plague, leaf plague, festival plague, neck plague and grain plague. The base of the affected rice seedling is grey or black, the upper end of the affected rice seedling turns into grey brown, and the affected rice seedling dies after rolling;
(4) Leaf spot of flax: the damage period is long, and diseases are possibly affected from rice seedlings to mature rice, the disease spots are obvious on leaves, and the disease spots are in brown shapes and are like round spots or elliptic spots of sesame. The seedlings are easy to die when being infected with diseases; when the mature rice is infected with diseases, the brown small spots gradually increase and take on a large ellipse shape. When the disease spots grow to the later stage, the middle piece is grey white, the periphery of the middle piece is brown, the outer ring is light yellow, and when the disease spots are more, the disease spots can be connected into one piece of spots.
(12) Carrying out data enhancement processing on the training set image by adopting image amplification and image contrast adjustment;
in practical problems, we need to identify pictures with unnecessary scenes and visual angles, in order to stably distinguish the images when the objects in the images are shifted, the visual angles of the objects are changed, the sizes of the images are changed, the brightness and illumination of the pictures are changed (or a combination of the images and the brightness and illumination of the images), and fine adjustment is carried out on the images in the image data set to obtain amplification data, so that the CNN model has invariance to the fine changes of some pictures and improves generalization capability of the model;
the contrast is poor if the contrast between the lesion and the background is small, and the contrast is large if the contrast between the lesion and the background is large. In consideration of the difficulty that the subsequent processing is not greatly increased if the contrast of the disease image is different, the contrast of the image needs to be enhanced, the details and the contours of the disease spots are enhanced, the image effect is improved, and the subsequent disease spot segmentation and feature extraction in a convolution network are easier to operate;
the image amplification method comprises the following steps:
changing the visual angle of the images in the image training set, firstly randomly selecting one of the images from the rotation angles of 0, 90, 180 and 270, and then randomly selecting and reversing the overturn of the input image to overturn along the x axis or along the y axis;
changing the size of the image, firstly cutting the image from the height and the width randomly by 0.1-0.4, and then carrying out random size scaling on the whole image by 0.2.
The image contrast adjustment in this embodiment adopts a color space-based image enhancement algorithm, specifically, adopts a uniform distribution for non-linearly converting the pixel value distribution of the original image into a first preset threshold interval, and adopts a non-linear transformation formula:
wherein G is out For the pixel value of each pixel after transformation, G in G is the original pixel value exmin And G exmax Lower and upper limits for a first preset threshold interval，G min And G max The average minimum pixel value and the maximum pixel value of the RGB color space are respectively calculated as follows:
G min ＝(G min ＝(R min +G min +B min )/3，G max ＝(R max +G max +B max )/3。
in this embodiment, the first preset threshold value is (90,255) adopted, so that a certain gray pixel value interval in which the pixel value distribution of the original image is relatively concentrated can cover the entire range of [0,255] as far as possible, by using this method, the R, G, and B3 color channels can be respectively subjected to automatic contrast stretching, and the obtained results are synthesized, thereby completing the enhanced image under the RGB color space.
(13) Obtaining a trained deep learning network, wherein the trained deep learning network is obtained through training of an image training set and a constructed deep learning network to be trained, and the construction and training of the deep learning network to be trained are realized based on an auxiliary model;
considering that the existing rice disease images do not have a huge number of data sets, training is directly performed by adopting a neural network, an ideal classification effect cannot be achieved, transfer learning is adopted, a network model based on large data set training is utilized, weight parameters and network model fine adjustment are performed, training is performed, training time is shortened, classification accuracy is improved, and in particular, in the embodiment, the auxiliary model selection method is that the task field of the auxiliary model is the same as the training task field of the image training set, the similarity between the data sets of the auxiliary model and the image training set meets a second preset threshold interval, namely [ alpha, beta ], the similarity is calculated by adopting a maximum mean difference algorithm MMD, when the MMD is smaller and smaller than an alpha value, the data sets of the auxiliary model are similar to the distribution of the image training set, image recognition can be directly performed by directly utilizing the auxiliary model with perfect training, generally, the source field and the target field distribution have a certain difference, namely the MMD is higher than a set threshold alpha, but the threshold beta is not exceeded, and a new hidden layer is required to be added or replaced by the network for the target field model to learn new knowledge. However, if the MMD is too large and exceeds the threshold value β, it indicates that the data distribution difference between the source domain and the target domain is too large, which may cause a negative migration phenomenon, and the task is not suitable for migration learning or needs to replace the source domain data model, where the threshold values α and β are set according to the results of multiple actual experiments.
In addition, since disease identification generally refers to identifying lesions of a leaf blade or other parts, a disease identification process based on machine learning generally includes: preprocessing an original image, dividing an image lesion, extracting lesion characteristics and identifying lesion types.
Aiming at rice disease images under a field complex background, each disease spot type in the rice disease images belongs to a fine granularity type, characteristic differences among the disease spots are not very large and are easily influenced by factors such as illumination, background and position, a set of more accurate and efficient image segmentation method needs to be researched, as the identification of the rice diseases is carried out in a complex field environment, crop disease symptoms are various, the situation that the rice images are overlapped among blades and damaged is likely to occur, and therefore, the image segmentation process is relatively difficult, the identification of the rice diseases is carried out by using a convolutional neural network algorithm, and the convolutional neural network does not need the processing process of image segmentation.
In the application, the convolutional neural network is also adopted for extracting the image features, various visual features of the image are extracted layer by layer through a CNN model and combined, CNN has stronger feature extraction capability, and the network layer in front of CNN can realize automatic extraction of shallow visual features, such as: the strong feature extraction capability of CNN is enhanced along with the increase of the number of layers, and finally abstract high-level representation combined features can be formed, so that the feature extraction of rice diseases is more accurate.
And the convolutional neural network CNN structure adopts a strategy of local connection and parameter sharing, so that the parameter number of the network can be effectively reduced, and the training speed of the network is greatly increased.
Based on the above, the type of the auxiliary model in the embodiment adopts the deep convolution network, so that the deep convolution network structure of the auxiliary model is migrated to the deep learning network to be trained, and image segmentation and feature extraction are performed.
The method for constructing and training the deep learning network to be trained based on the auxiliary model comprises the following steps:
(41) Acquiring a hidden layer and a hidden layer weight parameter of an auxiliary model;
(42) Migrating weight parameters of a convolution layer and a pooling layer of the auxiliary model and corresponding layers to a deep learning network to be trained;
(43) Adding a spatial pyramid pooling layer into a pooling layer in a deep learning network to be trained, fusing outputs of all the spatial pyramid pooling layers, and connecting to a full-connection layer, wherein the full-connection layer adopts 3 layers; the spatial pyramid pooling layers in the embodiment have different scales, so that the deep learning network to be trained can input images of any size and generate output of fixed size, the input images do not need to be subjected to size normalization processing, the feature loss caused by size normalization of the input images is avoided, the effectiveness of feature extraction is improved, and the accuracy of network model identification is improved.
The extracted image features have multiscale by adopting different sizes, so that the multiscale feature extraction of rice diseases is realized, and the disease identification accuracy is improved;
(44) The method comprises the steps of fixing weight parameters of a convolution layer and a pooling layer of a deep learning network to be trained, and inputting an image training set into the deep learning network to be trained to perform weight parameter training of a full-connection layer;
(45) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a first preset value, fixing the network parameters is canceled, and fine adjustment of the weight parameters of the whole network layer is performed;
(46) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than the second preset value or the iteration number reaches the maximum preset value, the training process is stopped, and the difference is gradually reduced in the training process.
The full-connection layer adopts 3 layers, L2 regularization is added in the original loss function of the full-connection layer in order to prevent overfitting, a Dropout method is adopted in each full-connection layer, in the training process, a part of neurons in an input layer or an hidden layer are randomly selected through predefined probability to reset the weight to 0, namely the part of neurons are discarded, and then non-discarded neurons are stretched. And then, continuously updating parameters according to the learning method of the neural network, and deleting some neurons at random in the next iteration until the training is finished. The difference of the labeling data corresponding to the output value and the input value of the network model is measured by adopting a cross entropy function, and on the basis of the difference, the difference measurement formula of the network model in the embodiment is regularization of the cross entropy function and L2.
In the training process, the weight parameters of each layer are reversely propagated and adjusted according to the cross entropy value at each time, and when the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a first preset value in the step (45), the fixation of the network parameters is canceled, the fine adjustment of the weight parameters of the whole network layer is carried out, the learning rate in the process is larger than that in the step (46), and an adam algorithm is adopted in the adjustment method for conveniently reasonably setting the two learning rates.
Aiming at the identification tasks of four diseases of bacterial leaf blight, false smut, rice blast and flax leaf spot in the embodiment, an output layer of 4 nodes is added at last.
And finally, (14) acquiring an image to be identified, preprocessing, inputting the image into a training-completed deep learning network, and acquiring an identification result of a target object in the image.
The present invention is not limited to the above-described specific embodiments, and various modifications may be made by those skilled in the art without inventive effort from the above-described concepts, and are within the scope of the present invention.
Claims (6)
1. The image recognition method based on deep learning is characterized by comprising the following steps of: comprising the following steps:
(11) Acquiring an image training set containing a target object, wherein images in the training set carry annotation data;
(12) Carrying out data enhancement processing on the training set image by adopting image amplification and image contrast adjustment;
the image amplification method comprises the following steps: changing the visual angle of the images in the image training set, firstly randomly selecting one of the images from the rotation angles of 0, 90, 180 and 270, and then randomly selecting and reversing the overturn of the input image to overturn along the x axis or along the y axis; changing the size of an image, firstly cutting the image randomly from the height and the width by 0.1-0.4, and then carrying out random size scaling on the whole image by 0.2;
the image contrast adjustment adopts a non-linear transformation formula which converts the pixel value distribution of an original image into uniform distribution of a first preset threshold interval:wherein G is out For the pixel value of each pixel after transformation, G in G is the original pixel value exmin And G exmax To be the lower limit and the upper limit of the first preset threshold interval, G min And G max The average minimum pixel value and the maximum pixel value of the RGB color space are respectively calculated as follows: g min ＝(G min ＝(R min +G min +B min )/3，G max ＝(R max +G max +B max )/3；
(13) Obtaining a trained deep learning network, wherein the trained deep learning network is obtained through training of an image training set and a constructed deep learning network to be trained, and the construction and training of the deep learning network to be trained are realized based on an auxiliary model;
the auxiliary model selecting method is that the task field of the auxiliary model is the same as the training task field of the image training set, and the similarity between the data set of the auxiliary model and the image training set meets a second preset threshold interval;
the construction and training of the deep learning network to be trained comprises the following steps: (41) Acquiring a hidden layer and a hidden layer weight parameter of an auxiliary model; (42) Migrating weight parameters of a convolution layer and a pooling layer of the auxiliary model and corresponding layers to a deep learning network to be trained; (43) Adding a spatial pyramid pooling layer into the pooling layer in the deep learning network to be trained, fusing the outputs of all the spatial pyramid pooling layers, and connecting to a full-connection layer; (44) The method comprises the steps of fixing weight parameters of a convolution layer and a pooling layer of a deep learning network to be trained, and inputting an image training set into the deep learning network to be trained to perform weight parameter training of a full-connection layer; (45) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a first preset value, fixing the network parameters is canceled, and fine adjustment of the weight parameters of the whole network layer is performed; (46) When the difference of the labeling data corresponding to the output value and the input value of the network model is smaller than a second preset value or the iteration number reaches a maximum preset value, stopping the training process;
(14) And acquiring an image to be identified, preprocessing, inputting the image into a training-completed deep learning network, and acquiring an identification result of a target object in the image.
2. The deep learning based image recognition method of claim 1, wherein: and the similarity between the data set of the auxiliary model and the image training set is acquired by adopting a maximum mean difference algorithm (MMD).
3. The deep learning based image recognition method of claim 2, wherein: the type of the auxiliary model adopts a deep convolution network.
4. A deep learning based image recognition method according to claim 3, wherein: the full-connection layer adopts 3 layers, and each full-connection layer adopts an L2 regularization and Dropout method.
5. A deep learning based image recognition method according to claim 3, wherein: the difference of the annotation data corresponding to the output value and the input value of the network model is measured by adopting a cross entropy function, and in the training process, the weight parameters of each layer are adjusted by carrying out counter propagation according to the cross entropy value at each output, and the adjustment method adopts an adam algorithm.
6. The application of the image recognition method based on deep learning according to any one of claims 1-5 in rice disease recognition, which is characterized in that: the images of the image training set comprise different kinds of disease images, images of the same kind of disease in different growth periods and images of the same kind of disease in different rice areas.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010561990.1A CN111833311B (en) | 2020-06-18 | 2020-06-18 | Image recognition method based on deep learning and application of image recognition method in rice disease recognition |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010561990.1A CN111833311B (en) | 2020-06-18 | 2020-06-18 | Image recognition method based on deep learning and application of image recognition method in rice disease recognition |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111833311A CN111833311A (en) | 2020-10-27 |
| CN111833311B true CN111833311B (en) | 2023-12-22 |
Family
ID=72897821
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010561990.1A Active CN111833311B (en) | 2020-06-18 | 2020-06-18 | Image recognition method based on deep learning and application of image recognition method in rice disease recognition |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111833311B (en) |
Families Citing this family (8)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112465820A (en) * | 2020-12-22 | 2021-03-09 | 中国科学院合肥物质科学研究院 | Semantic segmentation based rice disease detection method integrating global context information |
| CN113076873B (en) * | 2021-04-01 | 2022-02-22 | 重庆邮电大学 | Crop disease long-tail image identification method based on multi-stage training |
| CN113177578A (en) * | 2021-04-06 | 2021-07-27 | 安徽农业大学 | Agricultural product quality classification method based on LSTM |
| CN113095230A (en) * | 2021-04-14 | 2021-07-09 | 北京深睿博联科技有限责任公司 | Method and device for helping blind person to search for articles |
| CN113576508B (en) * | 2021-07-21 | 2024-07-09 | 华中科技大学 | Cerebral hemorrhage auxiliary diagnosis system based on neural network |
| CN113553985A (en) * | 2021-08-02 | 2021-10-26 | 中再云图技术有限公司 | High-altitude smoke detection and identification method based on artificial intelligence, storage device and server |
| CN114548265B (en) * | 2022-02-21 | 2024-06-21 | 安徽农业大学 | Crop leaf disease image generation model training method, crop leaf disease identification method, electronic equipment and storage medium |
| CN117011607A (en) * | 2023-08-08 | 2023-11-07 | 安徽农业大学 | Rice seed classification method based on attention residual error network |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273502A (en) * | 2017-06-19 | 2017-10-20 | 重庆邮电大学 | A kind of image geographical marking method learnt based on spatial cognition |
| CN108665447A (en) * | 2018-04-20 | 2018-10-16 | 浙江大学 | A kind of glaucoma image detecting method based on eye-ground photography deep learning |
| CN109635835A (en) * | 2018-11-08 | 2019-04-16 | 深圳蓝韵医学影像有限公司 | A kind of breast lesion method for detecting area based on deep learning and transfer learning |
| WO2019233166A1 (en) * | 2018-06-04 | 2019-12-12 | 杭州海康威视数字技术股份有限公司 | Surface defect detection method and apparatus, and electronic device |
| CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
-
2020
- 2020-06-18 CN CN202010561990.1A patent/CN111833311B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107273502A (en) * | 2017-06-19 | 2017-10-20 | 重庆邮电大学 | A kind of image geographical marking method learnt based on spatial cognition |
| CN108665447A (en) * | 2018-04-20 | 2018-10-16 | 浙江大学 | A kind of glaucoma image detecting method based on eye-ground photography deep learning |
| WO2019233166A1 (en) * | 2018-06-04 | 2019-12-12 | 杭州海康威视数字技术股份有限公司 | Surface defect detection method and apparatus, and electronic device |
| CN109635835A (en) * | 2018-11-08 | 2019-04-16 | 深圳蓝韵医学影像有限公司 | A kind of breast lesion method for detecting area based on deep learning and transfer learning |
| CN110781921A (en) * | 2019-09-25 | 2020-02-11 | 浙江农林大学 | Depth residual error network and transfer learning-based muscarinic image identification method and device |
Non-Patent Citations (1)
| Title |
|---|
| 基于迁移学习的卷积神经网络玉米病害图像识别;许景辉;邵明烨;王一琛;韩文霆;;农业机械学报(第02期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111833311A (en) | 2020-10-27 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111833311B (en) | Image recognition method based on deep learning and application of image recognition method in rice disease recognition | |
| Tm et al. | Tomato leaf disease detection using convolutional neural networks | |
| Jiang et al. | A tomato leaf diseases classification method based on deep learning | |
| CN109360206B (en) | Method for dividing rice ears in field based on deep learning | |
| Kour et al. | Particle swarm optimization based support vector machine (P-SVM) for the segmentation and classification of plants | |
| CN106778902B (en) | Dairy cow individual identification method based on deep convolutional neural network | |
| CN105718945B (en) | Apple picking robot night image recognition method based on watershed and neural network | |
| CN108108677A (en) | One kind is based on improved CNN facial expression recognizing methods | |
| CN106340016A (en) | DNA quantitative analysis method based on cell microscope image | |
| CN110610506A (en) | Image processing technology-based agaricus blazei murill fruiting body growth parameter detection method | |
| CN110503140B (en) | Deep migration learning and neighborhood noise reduction based classification method | |
| CN108596891A (en) | A kind of method of counting towards multiple types mixing silk cocoon | |
| Concepcion II et al. | Lettuce growth stage identification based on phytomorphological variations using coupled color superpixels and multifold watershed transformation. | |
| CN101980298A (en) | Multi-agent genetic clustering algorithm-based image segmentation method | |
| CN110827273A (en) | Tea disease detection method based on regional convolution neural network | |
| Sharma et al. | Detection of rice disease using bayes' classifier and minimum distance classifier | |
| CN111666946A (en) | Plant point cloud single-blade segmentation method based on point cloud over-segmentation and surface patch growth | |
| Wang et al. | Segmentation of corn leaf disease based on fully convolution neural network | |
| CN116012721B (en) | Deep learning-based rice leaf spot detection method | |
| CN111626969B (en) | Corn disease image processing method based on attention mechanism | |
| Dai et al. | PPLC-Net: Neural network-based plant disease identification model supported by weather data augmentation and multi-level attention mechanism | |
| CN114648806A (en) | Multi-mechanism self-adaptive fundus image segmentation method | |
| CN117495851A (en) | Image contour processing-based water environment microorganism detection method | |
| Lubis et al. | Classification of tomato leaf disease and combination extraction features using K-NN algorithm | |
| CN105631451A (en) | Plant leave identification method based on android system |
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 |