Method, medium and device for individual tracking and identification of scophthalmus maximus and breeding applications thereof

Abstract

The application provides a method for tracking and identifying individual turbot, a medium, a device and its breeding application, and belongs to the technical field of biological identification. The method is characterized in that: pictures of the ventral surface and the dorsal surface of turbot are collected, and the collected turbot pictures are pretreated; the PCA algorithm is used for dimension reduction on the abdominal image of the turbot sample; by setting the Euclidean distance constraint condition, the abdominal texture image of the turbot that does not meet the condition is removed, so that individual identification is realized; for the individual image that does not meet the condition and is removed, the fin texture is used as the basis, and a deep learning model algorithm is used to segment the fin on the image; based on the obtained image dataset of the fin, a deep learning network is designed, and the identification of all samples is completed. The application also provides a computer readable storage medium and a device for running the identification method, and an application of the method in breeding. The method has high overall recognition rate; one growth cycle of turbot is monitored, and the monitoring task from the seedling stage to the adult fish is completed.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of biological recognition, and specifically relates to a turbot individual tracking and identification method, medium, device and breeding application thereof. BACKGROUND

[0002] Turbot is a main fish species for factory breeding in northern China. With global warming, extreme environments occur frequently, and problems such as frequent diseases, poor stress resistance and quality decline are becoming increasingly serious. It is urgent to carry out breeding of traits such as stress resistance and disease resistance of turbot, and to improve the germplasm, which has important significance for promoting the development of marine fish seed industry. In fish breeding work, in order to improve the accuracy of breeding, electronic tags are often used. In the case of group culture, individuals can be tracked, and kinship files and trait performances can be established, so as to screen out excellent traits, design mating schemes, and establish core breeding populations. However, this marking method often causes trauma to the surface of the fish, is susceptible to bacterial infection, and causes death. In order to avoid damage, electronic tags cannot be used to track fish fry and juvenile fish, which affects the monitoring and evaluation of traits throughout the growth cycle and is a technical barrier to precision breeding. Moreover, the cost of electronic tags suitable for fish on the market is high, and the number of test groups is often large in individual trait tracking during breeding, so the cost of tags also restricts the number of test groups, resulting in reduced breeding accuracy. Therefore, it is a key technical breakthrough for precision breeding to develop non-invasive and low-cost individual identification technology based on the biological characteristics of fish.

[0003] With the development of computer vision technology, research on fish individual identification based on fish image samples has emerged. However, due to the limited identification features, the research is mostly based on fixed fish sample image sets for identification, which is not suitable for tracking individuals throughout the growth cycle, thereby limiting the application range of fish identification. SUMMARY

[0004] In view of the above technical problems, the present application provides a turbot individual tracking and identification method, medium, device and breeding application thereof. In the previous study, it was found that the fish belly pattern of turbot from juvenile fish (about 15g) to adult fish (about 500g) has individual differences and is unchangeable, and the texture of the fin (including dorsal fin, tail fin, belly fin and anal fin) also has individual differences. Although the color depth changes with growth, the texture remains basically unchanged. This provides a new idea and method for feature extraction of turbot individuals.

[0005] The present application solves the technical problems of the present application by the following technical solutions based on the above findings:

[0006] A turbot individual tracking and identification method, the method being as follows:

[0007] The first step is to collect the pictures of the ventral and dorsal surfaces of turbot and pre-process the collected turbot pictures, including mask annotation and image enhancement.

[0008] The second step is to reduce the dimension of the ventral image of the turbot sample by using the PCA algorithm, and to eliminate the turbot ventral texture images that do not meet the condition by setting the Euclidean distance constraint condition, so as to realize individual identification.

[0009] The third step is to segment the fin part (dorsal fin, tail fin, ventral fin and anal fin) on the image based on the fin texture of the individual image that does not meet the condition and is eliminated, by using the deep learning model algorithm.

[0010] The fourth step is to design a deep learning network based on the obtained fin image dataset to complete the individual identification of all samples.

[0011] As a more specific embodiment, in the second step, for the turbot individuals that meet the condition in the pre-processed turbot dataset, the dimension is reduced based on principal component analysis, and then the individual identification is completed by using a logistic regression algorithm, and the specific implementation process is as follows:

[0012] Step 2.1: For the continuously tracked turbot individuals, since the similarity of the ventral images is relatively high, there are only 5 effective ventral sample images for each individual at the same period. The image size is adjusted to 512*512 during training, and the image dataset is divided into training dataset and test dataset in the proportion of 4:1;

[0013] Step 2.2: For the samples in the training set, the training set images are converted into 1-dimensional 1-column data, and the average value of the matrix is calculated based on the column to obtain the average ventral image;

[0014] Step 2.3: The training image is unfolded to form a matrix, and the new data matrix is obtained by subtracting the matrix mean value, and the covariance matrix of the matrix is calculated to obtain its eigenvalue and eigenvector;

[0015] Step 2.4: The feature vectors are sorted based on the size of the eigenvalues to obtain the matrix of the feature vectors, and the dimension of the feature vectors is selected as 120 to obtain the corresponding projection matrix. The distance of the test picture in the projection matrix is calculated, and the turbot pictures with a distance value greater than 35000 are discarded. Finally, the test picture is obtained.

[0016] Step 2.5: The test picture uses a logistic regression algorithm to complete the classification identification, and the correct recognition rate is 100%, which completes the identification model verification.

[0017] As a more specific embodiment, in the third step, for the turbot individuals that do not meet the condition and are eliminated, the fin texture feature extraction is performed for identification, and the specific process is as follows:

[0018] Step 3.1: first, label the dorsal fin, caudal fin, anal fin or pelvic fin on the fin picture taken by using labelme software;

[0019] Step 3.2: divide the labeled fin image dataset into training set and test set, and divide each image into background and fin 2 parts;

[0020] Step 3.3: select the deep learning network of deeplabv3 of semantic segmentation, which includes resnet 50 backbone network, ASPP module (Atrous Spatial Pyramid Pooling Module) and FCN head part;

[0021] Step 3.4: input the image of the training set, adjust the image size to 2048*512, set the learning rate from 0.0001 to 0.01, and complete the segmentation of the fin image after 2000 rounds of training, with a segmentation accuracy of 93.99%;

[0022] As a more specific embodiment, in the fourth step, the individual identification model of Scophthalmus maximus based on fin (dorsal fin, caudal fin, anal fin and pelvic fin) image has the following specific process:

[0023] Step 4.1: based on the fin image dataset obtained by the segmentation algorithm, divide the training set and test set in a ratio of 4:1;

[0024] Step 4.2: adopt a deep learning network, the backbone network adopts resnet18 structure, the neck adopts a global pooling average module, the head adopts a linear classifier, and the classification function adopts softmax;

[0025] Step 4.3: the deep learning network optimizer selects the stochastic gradient descent algorithm, the cross entropy is the loss function, and the class is 40;

[0026] Step 4.4: after 200 rounds of training and learning, the weight of the deep learning network is obtained, and the individual identification accuracy is 92%.

[0027] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is suitable for being loaded and executed by a processor to perform a Scophthalmus maximus individual identification method.

[0028] The application also provides a computer device, which comprises a memory and a processor, the memory stores a computer program, and the computer program is executed by the processor to make the processor perform a Scophthalmus maximus individual identification method.

[0029] The application also provides application of the method in individual identification in turbot breeding, and the method can identify parent source information, genetic information and phenotype information of the whole growth cycle of the turbot individual; and is beneficial to screening of a core breeding population, breeding of a parent, development of a reasonable mating scheme, avoidance of inbreeding, protection of genetic diversity of germplasm and more accurate breeding.

[0030] Compared with the prior art, the application has the following beneficial effects:

[0031] 1. Based on the previous research, it is found that the fish belly pattern of turbot has individual difference and is unchangeable from juvenile fish (about 15g) to adult fish (about 500g), and the texture of the fin (including dorsal fin, tail fin, ventral fin and anal fin) also has individual difference, although the color depth changes with growth, but the texture is basically unchanged, based on this finding, a cascade feature-based turbot individual identification model is proposed, which has the advantages that: the traditional machine learning algorithm is used to identify the fish belly image which is easy to distinguish; the image segmentation model is used to extract the fin image of turbot, and the deep learning algorithm is used to extract the features, and the fin image with high complexity is identified; thereby the identification accuracy of the turbot individual is improved as a whole; the identification of multiple feature parts realizes the monitoring of the whole growth cycle of the turbot, and completes the monitoring task from the fry stage to the adult fish, which is innovative.

[0032] 2. The application is applied in turbot breeding, and can identify the parent source information, genetic information and phenotype information of the whole growth cycle of the turbot individual; is beneficial to screening of a core breeding population, development of a reasonable mating scheme, avoidance of inbreeding and protection of genetic diversity of germplasm; and since it is for individual whole growth cycle monitoring, the breeding is more accurate.

[0033] 3. The application can realize non-invasive tracking of the fish body, i.e. without using damaging methods such as electronic tagging and fluorescent marking to track the individual, thereby reducing the risk of pathogenic infection caused by surface damage of the fish body, saving energy and protecting the environment, and reducing the risk of aquaculture.

[0034] 4. The application can reduce the tracking cost of the turbot individual, and the electronic tag costs 12-15 yuan per tag according to the market price, and a batch of breeding trait tracking and monitoring generally has more than 2000 individuals, so the cost is greatly reduced. BRIEF DESCRIPTION OF DRAWINGS

[0035] Figure 1 is a turbot individual identification process;

[0036] Figure 2 is a data set of individual images of turbot at different growth stages;

[0037] Figure 3 is a turbot belly average image;

[0038] Figure 4 Segmentation method of turbot image data;

[0039] Figure 5 is a deepLabV3 network structure;

[0040] Figure 6 is the dorsal fin, tail fin, ventral fin and anal fin of the segmented turbot fish;

[0041] Figure 7 is a deep learning network structure for individual recognition based on turbot fin characteristics. DETAILED DESCRIPTION

[0042] The technical solutions of the present application will be further explained by examples below, but the protection scope of the present application is not limited in any form by the examples.

[0043] Example 1

[0044] In this example, 40 turbot fish were tracked and monitored for 2 years, and each individual was provided with a physical identification marker. Image samples of turbot fish were collected once every three months, and each time the abdominal and dorsal images were collected to construct the training data set used in the present application. Through continuous tracking, it was found that the abdominal pattern of turbot fish had obvious individual characteristics, so a turbot individual recognition method combining dorsal and abdominal characteristics was designed. The specific method is as follows:

[0045] A method for individual recognition of turbot, the specific process is as shown in Figure 1 The method steps are as follows:

[0046] Step 1: Use a smart phone to collect the abdominal and dorsal images of turbot, and pre-process the collected turbot images, mainly including mask annotation, image enhancement, etc. Figure 2 ).

[0047] Step 2: Use principle component analysis (PCA) to reduce the dimension of the abdominal image of the turbot sample, and by setting the Euclidean distance constraint condition, eliminate the abdominal texture images of turbot fish that do not meet the condition, to achieve an individual recognition accuracy of 100%.

[0048] Step 2.1: For the continuously tracked individual Scophthalmus maximus, due to the high similarity of the abdomen images taken, each individual has only 5 effective abdomen sample images in the same image, and a total of 200 images for 40 individuals. The image size is adjusted to 512*512 during training, and the image dataset is divided into training dataset and test dataset in the ratio of 4:1, that is, 5 images per individual, 4 images into the training dataset and 1 image into the test dataset.

[0049] Step 2.2: For the samples in the training set, they are converted into 1-dimensional 1-column data, and the 160 images in the training set are grouped into a 160-column matrix X = [x1 x2…x 160 ] in the matrix, where x i is a 512*512 = 262144-dimensional vector. The average value of the matrix can be calculated based on the following formula Thus, the average abdomen image of Scophthalmus maximus is obtained Figure 3 .

[0050]

[0051] Step 2.3: The training images are unfolded to form the matrix X minus the average matrix A new 262144*160 data matrix is obtained, and the covariance matrix is calculated, and its eigenvalues and eigenvectors are obtained, and the 160 eigenvalues are as follows:

[0052]

[0053] Step 2.4: Based on the size of the eigenvalue, the matrix of the eigenvector is obtained, the dimension of the eigenvector is selected as 120, the corresponding projection matrix is obtained, the distance of the test picture in the projection matrix is calculated, and the Cynoglossus test pictures with a distance value greater than 35000 are discarded, and finally 20 test pictures are obtained.

[0054] Step 2.5: The 20 test pictures obtained are identified by using the logistic regression algorithm, and the correct recognition rate is 100%, and the identification model verification is completed.

[0055] Step 3: For individuals who are not identified by fish abdomen images, the fin texture (dorsal fin, tail fin, ventral fin and anal fin) is used to segment the fin based on deep learning model algorithm.

[0056] Step 3.1: First, use the 5000 pictures taken to label the fin (dorsal fin, tail fin, ventral fin and anal fin) using labelme software.

[0057] Step 3.2: Divide the labeled image dataset into training set and test set, and divide each image into background and fin 2 parts, such asFigure 4 The closed area plotted with dots is the fin part, and the other is the background.

[0058] Step 3.3: Considering the sample size of the image and the complexity of the data, a deep learning network (DeepLabV3) for semantic segmentation is selected. Figure 5 The structure mainly includes a resnet 50 backbone network, an ASPP module (Atrous Spatial Pyramid Pooling Module), and an FCN head part.

[0059] Step 3.4: Input the image of the training set, adjust the image size to 2048*512, set the learning rate from 0.0001 to 0.01, and train the network for 2000 rounds to complete the segmentation of the fin part of the image. Figure 6 The segmentation accuracy is 93.99%.

[0060] Step 4: Based on the obtained image dataset of the fin part (dorsal fin, caudal fin, pelvic fin, and anal fin), a deep learning network is designed to complete individual recognition.

[0061] Step 4.1: Based on the fin image dataset obtained by the segmentation algorithm, the training set and the test set are divided according to the image quantity ratio of 4:1.

[0062] Step 4.2: Since only the fin part is in the fish sample image, the image complexity is not high, and therefore a deep learning network as shown in Figure 7 is designed. The backbone network of the network uses a resnet18 structure, the neck uses a global pooling average module, the head uses a linear classifier, and the classification function uses softmax.

[0063] Step 4.3: The deep learning network optimizer selects the stochastic gradient descent algorithm, the cross-entropy is the loss function, and the class is 40.

[0064] Step 4.4: After 200 rounds of training and learning, the weight of the deep learning network is obtained, the weight file size is 520M, and the test accuracy is 92%.

[0065] After the above method test, the individual correct recognition rate of 40 test pictures is (20*100%+20*92%) / 40=96%, so it can be seen that the accuracy of the individual recognition of Scophthalmus maximus proposed in the present application combined with the characteristics of the abdominal and caudal fin part has been improved compared with the single deep learning algorithm, and the complexity has also been reduced. The present application does not discuss how to efficiently implement the principal component analysis method or the deep learning method single method, because the performance improvement of the specific algorithm can be applied to the present application, thereby improving the overall recognition accuracy.

Claims

1. A method for tracking and identifying individual turbot, characterized in that, The method is as follows: The first step, collect the pictures of the ventral and dorsal surfaces of Scophthalmus maximus, and pretreat the collected pictures of Scophthalmus maximus; The second step, the PCA algorithm is used for dimension reduction for the ventral surface image of the Scophthalmus maximus sample, the big Scophthalmus ventral surface image that does not meet the condition is removed by setting the Euclidean distance constraint condition, and then the individual recognition is completed by using the logistic regression algorithm, specifically: Step 2.1: For the continuously tracked Scophthalmus individuals, due to the high similarity of the ventral surface images, 5 effective ventral surface sample images of each individual at the same period are taken; and the image data set is divided into training data set and test data set according to the proportion of 4:1; Step 2.2: For the samples of the training data set, convert them into 1-dimensional 1-column data, and form a matrix with the training data set images, calculate the average value of the matrix based on the column, and obtain the average ventral surface image; Step 2.3: The training image is unfolded to form a matrix, and the average value of the matrix is obtained by subtracting the average value of the matrix, and the covariance matrix of the matrix is calculated to obtain its eigenvalue and eigenvector; Step 2.4: Sort the eigenvalues based on their size to obtain the matrix of eigenvectors, select the dimension of the eigenvector as 120, obtain the corresponding projection matrix, calculate the distance of the test picture in the projection matrix, and discard the Scophthalmus ventral surface picture with a distance value greater than 35000, and obtain the final test picture; Step 2.5: The test picture uses the logistic regression algorithm to complete the identification of the category; The third step, for the individual image that cannot be identified by the ventral surface image, the fin texture on the dorsal surface image is used to segment the fin on the dorsal surface image based on the deep learning model algorithm; The specific process is as follows: Step 3.1: first, the fin is segmented by using the deep learning model algorithm, and the fin is labeled by using the labelme software, and the dorsal fin, tail fin, pelvic fin or anal fin is labeled on the fin image in the dorsal surface image; Step 3.2: divide the labeled fin image data set into training set and test set, divide each image into background and fin two parts, and the closed area drawn by points is the fin Step 3.3: select a deep learning network based on deeplabv3, and the structure of the deep learning network includes resnet50 backbone network, ASPP module and FCN head part; Step 3.4: input the image of the training set, and the deep learning network is trained for 2000 rounds to complete the segmentation of the fin image, and obtain the fin image data set; The fourth step, based on the obtained fin image data set, a deep learning network is designed to complete the individual identification.

2. The method of claim 1, wherein the individual flatfish is a turbot. The specific process of the fourth step is as follows: Step 4.1: based on the fin image data set obtained by the segmentation algorithm, divide the training set and test set according to the proportion of 4:1; Step 4.2: learn the training set by using the deep learning network, the backbone network of the deep learning network adopts the resnet18 structure, the neck adopts the global pooling average module, the head adopts the linear classifier, and the function of classification selects the softmax; Step 4.3: the optimizer of the deep learning network selects the stochastic gradient descent algorithm, the cross entropy is the loss function, and the class is 40; Step 4.4: The deep learning network is trained for 200 rounds to obtain the weight of the deep learning network.

3. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a computer program, and the computer program is suitable for being loaded and executed by the processor to perform the method for tracking and identifying individual turbot according to any one of claims 1-2.

4. A computer device, comprising: The device comprises a memory and a processor, and the memory stores a computer program, and the computer program is executed by the processor to make the processor perform the method for tracking and identifying individual turbot according to any one of claims 1-2.

5. Use of the method according to claim 1 for individual identification in breeding of Scophthalmus maximus, characterized in that, The method of claim 1 can identify the parent source information, genetic information and phenotype information of the whole growth cycle of the individual turbot, so as to screen the core breeding population, cultivate the parents, and make a reasonable mating scheme. The method of claim 1 can identify the parent source information, genetic information and phenotype information of the whole growth cycle of the individual turbot, so as to screen the core breeding population, cultivate the parents, and make a reasonable mating scheme.

Images