Fish intelligent recognition method and sorting system

Abstract

A fish intelligent recognition method and a sorting system. The method comprises the following steps: step 1, collecting a fish image data set; step 2, preprocessing the collected data set, comprising data normalization and data augmentation; step 3, building a ConvNeXt neural network; step 4, training a model; step 5, obtaining a fish image to be predicted; and step 6, inputting the image into the neural network to obtain a classification result, and outputting and displaying the classification result. The method achieves efficient fish feature recognition and recognition result display, thereby greatly reducing the recognition work of fishery workers, reducing the work intensity, and improving the work efficiency.

Description

Intelligent fish recognition method and sorting system TECHNICAL FIELD

[0001] The application belongs to the technical field of image recognition, and particularly relates to a fish image recognition technology field based on a neural network model. BACKGROUND

[0002] China's fishery production is huge, and the economic value is significant, accounting for a large proportion in international trade. Fish and its products are an important part of our diet due to their rich high-quality protein, unsaturated fatty acids, vitamins and minerals. In fishery farming and fishing, especially in large net fishing at sea, fish recognition has always been an extremely important work, but in the past, it has been extremely dependent on manpower, which is low in efficiency and has a poor working environment, and at the same time, the labor cost of this work is high.

[0003] Fishing by ship in the fishing ground is an important activity in fishery production, but in actual work, it is found that when the ship carries out fishing operation by fishing net, the fish products and related aquatic products obtained are various in species, and the economic values of various species of products are also different, thereby leading to the need for accurate identification and sorting and screening of the captured fish and related aquatic products in fishery production.

[0004] The existing fish catch sorting method is mostly based on weight and size for sorting, and cannot sort different fish species and the same species of fish with different qualities in the same batch of fish catch. At the same time, the traditional fish classification method mostly relies on color and shape feature extraction to establish a model manually, has high requirements for image pixels, is only for specific fields, and has poor general performance. The real-time intelligent sorting system for offshore fish catch using deep learning has high identification efficiency and strong general performance.

[0005] Patent publication No. CN112676027A "Ocean fishing fish catch sorting device" uses a mechanical separation device to separate fish, shrimp and shellfish caught by fish catch, but cannot accurately identify different types of fish, shrimp and shellfish, and subsequent accurate sorting still needs manual operation; patent publication No. CN110125025A discloses "fish automatic sorting system based on image recognition", which uses an image acquisition module, an image recognition module and a weighing module to realize fish sorting, needs to apply 5G communication technology, and is prone to communication problems and difficult to be widely promoted; patent publication No. CN211129430U discloses "a size fish classification device", which uses multiple channels with different widths to classify fish of different sizes, and this technical solution can only classify fish by size, cannot classify different fish species in the same batch of fish catch, and has poor applicability.

[0006] With the continuous development of artificial intelligence technology, the use of convolutional neural networks for image classification technology has gradually matured, and the real-time recognition and classification method for fish has great application value. At the same time, the traditional fish classification method relies on color and shape feature extraction to establish a model manually, which has high requirements for image pixels and is only suitable for specific fish species, and has poor general performance. While the convolutional neural network based on ConvNeXt is easy to build, and the pre-trained model has excellent transfer learning performance, which can be trained for different fish image classification requirements.

[0007] SUMMARY

[0008] The present application proposes a fish intelligent recognition method and sorting system to improve the robustness and accuracy of identification in practical applications.

[0009] In order to achieve the above invention purpose, the application adopts the following technical scheme:

[0010] A fish intelligent recognition method, characterized in that it comprises the following steps:

[0011] STEP 1, model building

[0012] A fish intelligent recognition model is built, which includes a species recognition model and a quality recognition model; the species recognition model and the quality recognition model are both ConvNeXt network models;

[0013] STEP 2, intelligent recognition

[0014] The image data to be identified is put into the fish intelligent recognition model, and the recognition result is output.

[0015] In order to realize the intelligent sorting of fish, a fish intelligent recognition model needs to be built first. We choose the ConvNeXt network model. ConvNeXt is a high-efficiency and high-performance deep learning model suitable for image classification tasks. After the model is built, we can input the image data to be identified into the model and output the recognition result. The fish intelligent recognition model contains a species recognition model and a quality recognition model. Through this method, we can realize the intelligent sorting of fish images.

[0016] Due to the existence of many interference factors in real applications, the limitations of convolutional neural networks in practical applications are amplified, making it impossible to achieve accurate recognition and classification. It is worth noting that the kind identification model and the quality identification model in the technical solution of the present application are both ConvNeXt network models; the ConvNeXt model is a deep learning model based on the Transformer architecture. Compared with the traditional convolutional neural network (CNN) architecture, the ConvNeXt model uses DepthwiseConvolution as a lightweight convolution operation. By splitting the standard convolution into DepthwiseConvolution and PointwiseConvolution two parts, the number of parameters is effectively reduced and the calculation efficiency is improved, so that a deeper and wider neural network can be constructed with smaller convolution kernels and fewer parameters. Moreover, the ConvNeXt model uses LayerNorm as part of the downsampling operation, that is, LayerNorm is applied after each downsampling to help reduce the differences between different samples, thereby improving training efficiency and final performance. In general, compared with other traditional CNNs, the ConvNeXt model has better robustness and universality in the actual fish sorting application field.

[0017] As a preferred embodiment of the present application, the identification result includes category data, size data and freshness data.

[0018] Category, size and freshness are important parameters for fish market transactions and processing. By using an automatic image processing system to obtain these data, work efficiency can be greatly improved and the uncertainty of manual measurement can be reduced.

[0019] As a preferred embodiment of the present application, the training step of the fish intelligent identification model is specifically:

[0020] STEP A, data preprocessing

[0021] Obtain fish image datasets of different species and different qualities, perform image preprocessing on the fish image datasets, and randomly classify to obtain a training set and a validation set;

[0022] STEP B, model training

[0023] Train the fish intelligent identification model based on the training set and the validation set.

[0024] STEP C, model verification:

[0025] Put the verification set into the trained fish intelligent recognition model for verification, and judge the performance of the fish intelligent recognition model based on the verification result and whether it needs to be continuously trained.

[0026] First, the fish image dataset is preprocessed, common image preprocessing includes scaling, cropping, normalization, etc., which can ensure that all images are on the same scale, improve the efficiency and accuracy of model training. In addition, noise and unclear images are removed to improve the overall quality of the data and ensure the effectiveness of model training. And the preprocessed image data is randomly classified to prevent model overfitting and ensure good performance on unseen data. By dividing the dataset into a training set and a validation set, cross-validation can be performed to improve the accuracy of training. In addition, the training set is used to train the model, and the validation set is used to adjust the model parameters and prevent overfitting to ensure that the model has good generalization ability. In short, such a design is conducive to building a robust and highly generalizable fish image sorting model that can more accurately and efficiently identify fish varieties and quality.

[0027] As a preferred embodiment of the present application, the image preprocessing in the data preprocessing of STEP A specifically includes one or more of picture rotation, contrast increase, random cropping, horizontal flipping, and Gaussian variation. Picture rotation, contrast increase, random cropping, horizontal flipping, and Gaussian variation are common means of existing image processing. Picture rotation: can make the target object in the image be detected and recognized at different angles, improving the adaptability of the model to different angles. Contrast increase: helps to highlight the target object in the image, improving its distinguishability in subsequent processing steps. Random cropping: can increase the diversity of training data, enabling the model to better generalize to actual application scenarios. Horizontal flipping: can make the target object in the image be detected and recognized in the horizontal direction, improving the adaptability of the model to horizontal direction changes. Gaussian variation: can be used to simulate the effects of images under different lighting conditions, improving the adaptability of the model to lighting changes. In summary, by performing multiple preprocessing operations on the image, the accuracy and robustness of image processing can be effectively improved, thereby improving the performance and effectiveness of the model in practical applications.

[0028] As a preferred embodiment of the present application, the training step of the fish intelligent recognition model further includes a model pre-training step:

[0029] The fish intelligent recognition model is trained using the ImageNet22K dataset to obtain a pre-trained model, and the network weights of the pre-trained model are fine-tuned using the ImageNet1K dataset to obtain a fine-tuned training model, and the training set and the validation set are trained based on the fine-tuned training model.

[0030] ImageNet22K is a large visual database that contains 22,000 categories, covering a wide range of objects, scenes, and concepts. This database was created for computer vision research, particularly in the fields of machine learning and artificial intelligence, to train and evaluate the capabilities of image recognition systems. ImageNet1K is a subset of the ImageNet database consisting of 1,000 categories. The strategy of first training on a large dataset and then fine-tuning on a small dataset can help fish intelligence recognition models achieve better performance and generalization ability, while improving training efficiency, in line with the goal of developing efficient and accurate artificial intelligence models.

[0031] An offshore fish catch intelligent sorting system, characterized by comprising a transmission device module, an image acquisition and recognition module, and a sorting device module, the image acquisition and recognition module includes an optical image acquisition device located above the transmission device module, and an image recognition device electrically connected with the optical image acquisition device; the sorting device module is provided with multiple groups, each group includes a cylinder electrically connected with the image recognition device and a corresponding target fish bin; the optical image acquisition device is used to acquire the target object image located on the transmission device module, the image recognition device is used to execute the above-mentioned fish intelligence recognition method, to identify and analyze the target object image, to determine the corresponding target fish bin, and to drive the cylinder in the same group to sort the target object into the target fish bin at the corresponding position.

[0032] The fish and other aquatic products to be identified are placed on the transmission device module for transmission, the optical image acquisition device located above the transmission device module acquires the images of the fish and other aquatic products to be identified on the transmission device module, and transmits them to the image recognition module for identification. The optical image acquisition device can be an industrial camera such as CCD, CMOS, etc. in existing technology, which can be used for long-time work in continuous operation of industrial environment, with good shock resistance and durability. Moreover, the industrial camera also has high resolution, which can capture clearer images, especially important for capturing details, which is conducive to accurate identification of the target object image by the image recognition module. The image recognition module is used to execute the above-mentioned fish intelligence recognition method, to process and analyze the acquired target object image, to determine the corresponding target fish bin, and to drive the cylinder in the corresponding target fish bin to move, so as to sort the target object into the target fish bin at the corresponding position.

[0033] The technical scheme of the present application realizes automatic sorting through the cooperation of the transmission device module, the image acquisition and recognition module, and the sorting device module, improves the sorting efficiency, and reduces the labor cost. In addition, the system can realize real-time image acquisition and recognition of the products on the transmission device, make quick sorting decisions, and maintain the continuity of the production line. In general, this design can realize efficient, accurate and automatic identification and sorting of fish and other aquatic products, which is of great significance to improving the automation level and product quality of the aquatic product processing industry.

[0034] As a preferred embodiment of the present application, a plurality of optical image acquisition devices are provided between adjacent two groups of sorting device modules.

[0035] An optical image acquisition device is provided between adjacent two groups of sorting device modules, so that each group of sorting device modules is equipped with an image acquisition module. This design can ensure that the same target object is identified and distinguished once before each target fish bin, which can ensure that each part of the target object is fully identified and avoid identification errors or missed sorting due to insufficient single angle or view.

[0036] As a preferred embodiment of the present application, the optical image acquisition device comprises a light-shielded chamber on the transmission belt, a light-emitting device in the light-shielded chamber, and a camera device.

[0037] The design of the light-shielded chamber can effectively isolate the influence of external light on image acquisition, ensuring the clarity and accuracy of the image. This is particularly important in the application scenario of fish identification and sorting, because changes in external light may cause image quality to decline, affecting the accuracy of identification. In addition, by using a light-emitting device in the light-shielded chamber, uniform illumination can be provided for the target object to be identified, which helps to eliminate problems such as shadows and light reflection caused by uneven illumination, thereby further improving the quality of the image and the accuracy of the identification. At the same time, since the light-shielded chamber design places the light-emitting device and the camera device in a closed and relatively stable environment, it also prolongs the service life of the device and improves the stability of the system.

[0038] As a preferred embodiment of the present application, the light-shielded chamber comprises a top light-shielded plate parallel to the transmission belt and a peripheral light-shielded plate perpendicular to the transmission belt, the light-emitting device is provided with a plurality of symmetrically arranged on the peripheral light-shielded plate, and the camera device is arranged on the top light-shielded plate.

[0039] The plurality of light emitting devices are symmetrically arranged on the peripheral light shielding plate, which can ensure that the target object is uniformly illuminated when passing through the light shielding chamber. Such uniform illumination helps to improve image quality, reduce shadows and reflections caused by uneven illumination, and further improve recognition accuracy.

[0040] As a preferred embodiment of the present application, the inner surface of the light shielding chamber is coated with a diffuse reflection layer.

[0041] The diffuse reflection layer can be selected from existing white paint with particles or powder, which can achieve diffuse reflection and has a long service life. In addition, the diffuse reflection layer can uniformly disperse the incident light into the entire light shielding chamber, avoiding local over-brightness or over-darkness, and ensuring that the target object is uniformly illuminated at all angles. Uniformly distributed light helps to reduce halos and spots in the image, making the image clearer and more detailed, and further improving the quality and accuracy of image recognition and processing.

[0042] As a preferred embodiment of the present application, the transmission device module includes a transmission belt and a blocking device above the transmission belt, and a fish passing gap is provided between the blocking device and the transmission belt.

[0043] When multiple fish and other aquatic products are stacked together, the blocking device can separate the upper stack, and the fish passing gap and the blocking device cooperate to ensure that the aquatic products passing through the blocking device are not stacked, reducing confusion caused by stacking during image recognition and improving the accuracy of subsequent image recognition.

[0044] As a preferred embodiment of the present application, the blocking device includes a shaft support frame connected to the transmission belt and an elastic lever assembly sleeved on the shaft support frame.

[0045] The elastic lever assembly can be selected from a plurality of blocking strips made of rubber material in the prior art. When multiple fish and other aquatic products on the transmission belt are stacked together, the upper products can be separated from the lower fish products while being subjected to resistance from the blocking device, and the surface of the fish products can be prevented from being scratched by the blocking device, thereby affecting the quality of the fish. In addition, the elastic lever assembly rotates on the shaft support frame, and the elastic lever assembly itself can deform under stress, which can prevent the transmission process from being blocked due to the large size of individual fish products, and ensure stable operation of the system.

[0046] As a preferred embodiment of the present application, the shaft support frame is provided with a sliding guide rail, which realizes sliding perpendicular to the transmission direction of the transmission belt.

[0047] The size of the fish passing gap can be flexibly adjusted by matching the sliding guide rail and the shaft support frame, so as to adapt to different sizes of products to be sorted, and the adaptability of the offshore fish catch intelligent sorting system is improved.

[0048] In summary, the present application has the following advantages:

[0049] 1. Different fish species and qualities require different features to distinguish, and using multiple models allows each model to focus on a specific sorting task, thereby better capturing and utilizing relevant features.

[0050] 2. The step-by-step training and verification process helps build a high-performance fish intelligent recognition model that can accurately identify and classify different types of fish, thereby improving sorting efficiency and accuracy.

[0051] 3. The cooperation between the transmission device module, image acquisition and recognition module, and sorting device module realizes automatic sorting, improves sorting efficiency, and reduces labor costs. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0053] Figure 1 is a flow chart of a fish intelligent recognition method based on ConvNeX;

[0054] Figure 2 is a flow chart of the training steps of the fish intelligent recognition model;

[0055] Figure 3 is a schematic diagram of the fish intelligent recognition model device;

[0056] Figure 4 is a schematic diagram of an optical image acquisition device;

[0057] Figure 5 is a schematic diagram of a transmission device module.

[0058] In the figure, 1 is a transmission device module, 11 is a transmission belt, 12 is a blocking device, 121 is a shaft support frame, 122 is an elastic lever assembly, 2 is an image acquisition and recognition module, 21 is an optical image acquisition device, 211 is a light shielding chamber, 212 is a light emitting device, 213 is a camera device, 22 is an image recognition device, 3 is a sorting device module, 31 is a cylinder, 32 is a target fish bin. DETAILED DESCRIPTION

[0059] The application will be further described below with reference to specific embodiments and drawings. The embodiments of the application are intended to better enable those skilled in the art to understand the application and are not intended to limit the application in any way.

[0060] Embodiment one

[0061] As shown in FIG. 1, the fish intelligent recognition method based on ConvNeX is as follows:

[0062] STEP 1, model building

[0063] In order to realize intelligent sorting of fish, a fish intelligent recognition model is built, which includes a species recognition model and a quality recognition model; both the species recognition model and the quality recognition model are ConvNeXt network models; the model structure includes 4 branches, each branch contains a 3x3 depth separable convolution layer in series, followed by a 1x1 convolution layer and a GELU activation function with a discard function, then the outputs of each branch are combined through a global average pooling layer, and classified through the last layer of full connection layer with the number of nodes of fish image species

[0064] The ConvNeXt convolutional neural network structure refers to the Transformer architecture for convolutional network design, adopts Depthwise Conv as the convolution layer, adopts GELU as the activation function, and adopts LayerNorm as the down-sampling function;

[0065] The activation function GELU is as follows:

[0066] The down-sampling function LayerNorm is as follows:

[0067] (wherein:

[0068] -LayerNorm(x) represents the result of Layer Normalization on the input x.

[0069] -γ and β are learnable parameter vectors, both of which have a dimension of d. They are used to scale and shift the normalized results to allow the network to freely learn to adjust the mean and variance of the data.

[0070] -μ is the mean of the input x in the feature dimension, and the calculation method is:

[0071] -σ is the standard deviation of the input x in the feature dimension, and the calculation method is:

[0072] - e is a small constant (e.g., 10 -5 ), used to avoid division by zero.

[0073] As shown in FIG. 2, the training steps of the fish intelligent recognition model are as follows:

[0074] STEP A, data preprocessing

[0075] Different kinds and different quality fish image data sets are obtained, and the collected picture data is preprocessed. First, the picture resolution is uniformly scaled to 224*224. And data augmentation, including picture rotation, increasing contrast, random cropping, horizontal flipping, Gaussian change. The augmented data set is randomly divided into training set and data set in the ratio of 8:2.

[0076] STEP B, model training

[0077] The fish intelligent recognition model is trained based on the training set and the validation set. Specifically, the network weight is iterated multiple times through multiple Epochs, the weight is saved once for each iteration, and after multiple iterations, the model with the highest accuracy is selected as the trained recognition model.

[0078] STEP C, model verification:

[0079] The validation set is put into the trained recognition model for verification, and the performance of the fish intelligent recognition model is judged based on the verification result whether it needs to be trained.

[0080] The technical scheme of the present application is very suitable for large-scale data application, which can reduce the amount of calculation and more targeted quality discrimination, and improve the overall sorting accuracy. In addition, different fish and quality may need different features to distinguish, using multiple models allows each model to focus on a specific sorting task, so as to better capture and use relevant features.

[0081] Example two

[0082] As another embodiment of the present application, before STEP B, model training, the fish intelligent recognition model is first trained using the ImageNet22K dataset to obtain a pre-trained model, then the network weights of the pre-trained model are fine-tuned using the ImageNet1K dataset to obtain a fine-tuned training model, and the training set and the validation set are trained based on the fine-tuned training model. ImageNet22K is a large visual database that contains 22,000 categories covering a wide range of objects, scenes and concepts. This database was created for computer vision research, especially in the fields of machine learning and artificial intelligence, to train and evaluate the capabilities of image recognition systems. ImageNet1K is an image dataset consisting of 1,000 categories, which is a subset of the ImageNet database. The strategy of first training on a large dataset and then fine-tuning on a small dataset can help the fish intelligent recognition model achieve better performance and generalization ability, while improving training efficiency, in line with the goal of developing efficient and accurate artificial intelligence models.

[0083] Embodiment three

[0084] A fish intelligent recognition sorting system

[0085] As shown in FIG. 3, the fish and other aquatic products to be identified are placed on the conveying device module 1 for conveying, and the optical image acquisition device 21 located above the conveying device module 1 acquires images of the fish and other aquatic products to be identified on the conveying device module 1 and transmits them to the image recognition module for identification.

[0086] As shown in FIG. 4, the optical image acquisition device 21 comprises a light-shielded chamber 211 on the conveying belt 11, light-emitting devices 212 in the light-shielded chamber 211, and a camera 213. The light-emitting devices 212 are symmetrically arranged on the peripheral light-shielded plate, and the camera 213 is arranged on the top light-shielded plate. The inner surface of the light-shielded chamber 211 is coated with a diffuse reflection layer. The light-emitting devices 212 can be commonly used LEDs on the market, which have the characteristics of small size, long service life, and high luminous efficiency, or controllable cost incandescent lamps. The camera 213 can be an industrial camera in the prior art, such as CCD, CMOS, etc., which can be used for long-time work in a continuous industrial environment, has good shock resistance and durability. Moreover, the industrial camera also has high resolution, can capture clearer images, and is particularly important for capturing details, which is conducive to accurate identification of the target object image by the image recognition module. The image recognition module can be a computer in the prior art installed with fish recognition software, which inputs the image data acquired by the optical image acquisition device 21 into the image recognition device 22 for image recognition. The image recognition device 22 is used to execute the fish intelligent recognition method in embodiment one or embodiment two, identifies the corresponding target fish bin 32 after processing and analyzing the acquired target object image, and drives the corresponding target fish bin 32 and the air cylinder 31 in the same group to move, so that the target object is sorted into the corresponding target fish bin 32.

[0087] Embodiment four

[0088] In another embodiment of the present application, a corresponding optical image acquisition device 21 is arranged in front of each group of sorting device modules 3, so that the same target object to be identified is optically imaged and identified once before each target fish bin 32. In this way, each part of the target object can be fully identified, and identification errors or missed sorting caused by a single angle or insufficient view can be avoided.

[0089] As shown in FIG. 5, in another embodiment of the present application, the conveying device module 1 is specifically a conveying belt 11 and a blocking device 12 above the conveying belt 11. The blocking device 12 comprises a rotating shaft support frame 121 connected with the conveying belt 11 and an elastic lever assembly 122 sleeved on the rotating shaft support frame 121. The elastic lever assembly 122 is composed of a plurality of rubber strips with the same length, and a circular through hole is arranged at the center of each rubber strip to rotate on the rotating shaft support frame 121. A fish passing gap is arranged between the elastic lever assembly 122 and the conveying belt 11. When a plurality of fish and other aquatic products are stacked together, the product on the top can be separated from the fish below after being subjected to the resistance of the rubber strip. In addition, the elastic lever assembly 122 rotates on the rotating shaft support frame 121, and the elastic lever assembly 122 itself can be deformed under force, which can avoid the blockage in the conveying process caused by the large size of a single fish product, ensure the stable operation of the system, and avoid the scratching of the fish surface by the blocking device 12 to affect the quality of the fish.

[0090] In another embodiment of the present application, a sliding guide rail is arranged on the rotating shaft support frame 121, which can be a ball friction guide rail or a pneumatic guide rail in the prior art. The sliding guide rail realizes the sliding perpendicular to the conveying direction of the conveying belt 11, which can flexibly adjust the size of the fish passing gap to adapt to the products with different sizes for sorting, and improves the adaptability of the offshore catch intelligent sorting system.

[0091] It should be noted that the above embodiments only explain the preferred embodiments and principles of the present application, and for ordinary skilled persons in the art, the specific embodiments can be changed according to the idea provided by the present application, and these changes are also considered as the protection scope of the present application.

Claims

1. A fish intelligent identification method, characterized in that, Comprising the following steps: STEP 1, model building Build a fish intelligent identification model, which includes a species identification model and a quality identification model; the species identification model and the quality identification model are both ConvNeXt network models; STEP 2, intelligent identification Put the image data to be identified into the fish intelligent identification model, and output the identification result.

2. The fish intelligent recognition method according to claim 1, characterized in that, The identification result includes category data, size data and freshness data.

3. The fish intelligent recognition method according to claim 2, characterized in that, The training steps of the fish intelligent identification model are as follows: STEP A, data preprocessing Obtain fish image data sets of different species and different qualities, perform image preprocessing on the fish image data sets, and randomly classify to obtain a training set and a validation set; STEP B, model training Train the fish intelligent identification model based on the training set and the validation set. STEP C, model verification: Put the validation set into the trained fish intelligent identification model for verification, and judge the performance of the fish intelligent identification model based on the verification result and whether it needs to be trained further.

4. The fish intelligent recognition method according to claim 3, characterized in that, The image preprocessing in the data preprocessing of STEP A specifically includes one or more of picture rotation, contrast increase, random cropping, horizontal flipping, and Gaussian variation.

5. The fish intelligent recognition method according to claim 3, characterized in that, The training steps of the fish intelligent identification model also include a model pre-training step: Train the fish intelligent identification model using the ImageNet22K data set to obtain a pre-trained model, then fine-tune the network weights of the pre-trained model using the ImageNet1K data set to obtain a fine-tuned training model, and train the training set and the validation set based on the fine-tuned training model.

6. An intelligent fish recognition and sorting system, comprising a conveying device module (1), an image acquisition and recognition module (2), and a sorting device module (3), wherein the image acquisition and recognition module (2) comprises an optical image acquisition device (21) located above the conveying device module (1), and an image recognition device (22) electrically connected to the optical image acquisition device (21). The image recognition device (22) is used to execute a fish intelligent identification method as claimed in any one of claims 1-6.

7. The fish intelligent recognition and sorting system according to claim 7, characterized in that, The transmission device module (1) includes a transmission belt (11) and a blocking device (12) located above the transmission belt (11), and a fish passing gap is provided between the blocking device (12) and the transmission belt (11).

8. The fish intelligent recognition and sorting system according to claim 7, characterized in that, The blocking device (12) includes a rotating shaft support frame (121) connected with the transmission belt (11) and an elastic lever assembly (122) sleeved on the rotating shaft support frame (121).

Images