BE1029938B1 - METHOD, PLATFORM AND TERMINAL FOR FISH IDENTIFICATION IN FISH FISH FARMING - Google Patents

Abstract

A method for identifying fish in fish farming is disclosed, comprising: developing a bank of fish image samples; extracting shapes and positions of fish heads and fish tails from fish image samples to construct a joint fish identification feature bank of fish bodies, fish heads and fish tails; developing a fish identification transfer learning model, pre-training the fish image sample bank according to the fish identification feature bank, and identifying fish in fish farming through the transfer learning model of fish identification; constructing a small data set of sample samples to test; training the pre-learned fish identification transfer model again using the small sample set; and inputting the resulting fish images into the fish identification transfer model to identify individual fish and fish species.

Description

METHOD, PLATFORM AND TERMINAL FOR FISH IDENTIFICATION IN

FISH CULTURE

BACKGROUND OF THE INVENTION

[0001] 1. Technical field

[0002] The present invention belongs to the technical field of fish farming, in particular to a method, a platform and a terminal for identifying fish in fish farming.

[0003] 2. Description of the related art

[0004] In the process of fish production and environmental protection monitoring, it is very important to guide aquaculture production by accurately measuring the phenotypic statuses of fish underwater, and then obtaining the sizes fish and accurately estimating their growth stages.

To achieve effective measurement on underwater fish phenotypes, accurate and efficient identification of underwater fish is a prerequisite. In addition, if fish body shape directions can be identified on the basis of fish identification, the search range of characteristic points of phenotypic measurements will be effectively reduced, and a foundation will be laid for realizing high-speed measurement. flow rate of underwater targets.

BRIEF SUMMARY OF THE INVENTION

[0005] One embodiment of the present invention relates to an integrated method of identifying individual fish underwater and body shape directions based on transfer learning, comprising the steps of:

[0006] development of a bank of fish image samples;

[0007] extraction of shapes and positions of fish heads and tails

° BE2023/5222 fish from samples of fish images to develop a joint fish identification feature bank of fish bodies, fish heads and fish tails;

[0008] development of a transfer learning model for fish identification, pre-learning of the bank of fish image samples according to the bank of fish identification characteristics, and identification of fish in fish farming through the fish identification transfer learning model;

[0009] construction of a small set of samples of samples to be tested, where the small set of samples refers to a set of a relatively small number of samples in a specific scenario;

[0010] training the pre-learned fish identification transfer model again using the small set of samples; And

[0011] input of the obtained fish images into the fish identification transfer model to identify individual fish.

BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0012] On reading the following detailed description with reference to the accompanying drawings, the above objectives, characteristics and advantages and other objectives, characteristics and advantages of the exemplary embodiments of the present invention will be easier to understand. In the drawings, a number of embodiments of the present invention are shown in a manner given by way of example and not limitation.

[0013] Fig. 1 is a flowchart of a method for identifying fish in fish farming according to one embodiment of the present invention.

[0014] Fig. 2 is a structural schematic diagram of a network model

> BE2023/5222 for identification of CNNF-1 fish bodies according to one embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

[0015] At present, identification technology for underwater fish targets mainly employs an identification method for a specific set of samples. When the changes of identified targets exceed an initial sample set space, it is necessary to create a new sample set and reconstruct an identification algorithm, which is time-consuming and laborious and not conducive to popularization and application of technology. Particularly when there are few observed fish samples, existing identification methods based on a specific sample set cannot achieve effective identification. Furthermore, existing learning algorithms only identify individual fish, and there is no identification method that integrates individual identification with body shape direction identification.

[0016] As an important incremental learning method, transfer learning has been widely used in many fields, but there is still no effective fish identification model. In addition, there are few fish samples in the international general sample bank at present, which cannot meet the needs of fish transfer learning.

[0017] According to one or more embodiments, a method for identifying fish in fish farming, as shown in Fig. 1, includes the following steps:

[0018] S101, typical fish with representative significance are selected to develop a fish sample bank, and images are preprocessed according to unified standards;

* BE2023/5222

[0019] S102, the positions of fish heads and fish tails are extracted from preprocessed data to develop a bank of joint characteristics of fish bodies, fish heads and fish tails;

[0020] S103, the preprocessed data is classified to obtain a training set and a test set;

[0021] S104, a fish transfer learning model is developed and pre-learned with a bank of classification samples;

[0022] S105, a small data set of samples of samples to be tested is constructed;

[0023] S106, the pre-learned transfer model is re-trained with small samples;

[0024] S107, fish images are input to the trained model to identify individual fish first;

[0025] S108, a minimum bounding rectangle of the individual fish is developed, and then the horizontal axis and the vertical axis of the bounding rectangle are expanded by 10 pixels respectively to develop a new identification zone, in which the positions of fish heads and fish tails are identified using a transfer learning algorithm;

[0026] S109, integrated positioning of individual fish and body shape direction is implemented by taking the center point of the minimum bounding rectangle of fish heads and fish tails as the directional positioning point, and the direction of a line connecting fish tails to fish heads as the direction of body shape.

[0027] The embodiment of the present invention has the beneficial effect that the efficiency and accuracy of fish image identification in an existing fish farm are improved by the integrated fish identification method and

> BE2023/5222 for body shape direction identification based on transfer learning.

[0028] According to one or more embodiments, a method for identifying fish in fish farming comprises the following steps:

[0029] S1, for common domestic aquaculture species and protected species, common cypriniformes and sliuriformes fish, such as silver carp, cyprinoid, crucian carp, macrura reevesii, pelteobagrus fulvidraco, acipenser dabryanus and leiocassis longirostris, from from which a large number of samples are easily obtained, are previously classified into cyprinoideae, hypophthalmichthyinae, cobitinae, acipenserinae and bagridae to develop a bank of fish samples of families and genera.

[0030] S2, training and test sample sets are constructed based on fish body, fish head and fish tail targets, and images in the sample sets are calibrated according to mirror images by rotating -30°, -60°, 30° and 60° and scaling 1:2, 1:0.75, 1:0.5 and 1:0.25 respectively to increase the number of samples.

[0031] S3, a fully connected layer is added at the end of a pre-trained model VGG19 to develop a transfer learning model.

[0032] S4, a batch gradient descent optimization algorithm in which BatchSize is defined as 20 is selected, a random gradient descent algorithm in which the learning rate is defined as 0.0001 is selected , and Softmax is selected as the loss function.

[0033] S5, the parameters of all convolutional layers outside the fully connected layer are frozen, and the developed model is trained with subfamily classes as output classes of the fully connected layer and the bank of fish samples of families and genera as a basis for obtaining a preliminary fish classification convolutional model package.

© BE2023/5222

[0034] S6, target object data of a specific crossing scenario or a biological investigation scenario are collected, and a small bank of samples to identify the targets accurately is developed according to the targets to be identified .

[0035] S7, the trained transfer learning model is retrained by opening the parameters of the fully connected layer and the 1 to 3 adjacent convolutional layers.

[0036] S8, a small batch gradient descent optimization algorithm in which MiniBatchSize is set to be 10 is selected, a random gradient descent algorithm in which the learning rate is set to be 0.0001 is selected, Softmax is selected as the loss function, and the data is reshuffled once per training cycle.

[0037] S9, the experimental results are compared and analyzed by integrating the identification accuracy and the identification efficiency, and network models are determined when a few convolutional layers adjacent to the fully connected layer are selected and the identification accuracy error between two selected convolutional layers is less than 1%. A fish body identification network CNNF-1, a fish head identification network

CNNFH-2 and CNNFT-2 fishtail identification network are constructed respectively.

[0038] S10, images of fish in a real working environment are collected, and RGB channels of the initial images are filtered using Gaussian filtering to eliminate image noise.

[0039] S11, the actual sampled data is identified using the trained transfer learning model, where the positions of individual fish in the data are first identified using the CNNF-1 model, and a

7 BE2023/5222 minimum bounding rectangle of identified targets is determined.

[0040] S12, taking the center of the rectangle as the positioning reference point, the length and width directions of the minimum bounding rectangle undergo an expansion of 10 pixels respectively to increase the length by 20 pixels and the width by 20 pixels in total, to construct a new area to be identified.

[0041] S13, the positions of fish heads and fish tails are identified in the new area to be identified, using the models CNNFH-2 and

CNNFT-2 respectively.

[0042] S14, when the positions of fish heads and fish tails can be identified simultaneously, the identified fish objects are retained to determine the movement directions in the next step; or when the positions of fish heads and fish tails cannot be identified simultaneously, the targets are dropped, but the targets are stored as replacement sample bank images to reserve the samples for incremental learning.

[0043] S15, a minimum bounding rectangle of the identified positions of fish heads and fish tails is constructed, a central position of the rectangle is calculated, the fish tails and the fish heads are connected by straight lines, and the directions of the straight lines are taken as the target movement directions of the fish.

[0044] According to one or more embodiments, experimental samples are collected with common cultivated fish and protected fish species cultivated in a protection center as targets, where the pre-learned model collects 12330 images of grass carp, cyprinoid, macrura reevesii, spotted silver carp, silver carp, acipenser dabryanus, etc. Fish to test,

selected from the implementation base, are crucian carp, parabramis pekinensis and leiocassis longirostris, and 200 images are collected from each fish. The pre-trained model is trained with small samples to build the final identification model.

[0045] During implementation in the field, 280 images of crucian carp, 220 images of parabramis pekinensis, and 86 images of leiocassis longirostris are collected. The fish identification results show that the identification rate of crucian carp is 99%, the identification rate of parabramis pekinensis is 100%, the identification rate of leiocassis longirostris is 98.8%, and the overall fish identification accuracy reaches 99.6%.

[0046] Fish heads and fish tails are positioned and identified from the identified fish images, and the identification accuracies are for crucian carp (fish heads: 99%, fish tails: 98%), parabramis pekinensis (fish heads: 98%, fish tails: 98%), and leiocassis longirostris (fish heads: 100%, fish tails: 97%). The structure of the developed fish identification transfer learning model is as follows.

[0047]

Parameter Layer Name Network Parameter Layer Name

ImageinputImageinput Size: (224, 224, 3) Conv512 2 Size: (3, 3)

Number: 512

Convé4_1 Size: (3, 3) Number: 64 | Conv512 3 Size: (3, 3)

Number: 512

Conv64 2 Size: (3,3) Number: | Maxpool512 2 Size: (3, 3) 64 Step length: (2,2

Maxpool64_1 Size: (3, 3) Conv512_4 Size: (3, 3)

Number Length: 512 Step: (2.2

Conv128 1 Size: (3.3) Number: | Conv512 5 Size: (3, 3) 128 Number: 512

Conv128 2 Size: (3.3) Number: | Conv512 6 Size: (3, 3) 128 Number: 512

Maxpool128 1 Size: (3,3) Maxpool512 3 Size: (3, 3)

Step Length Length: (2,2 Step Length: (2, 2

Conv256 1 Size: (3.3) _ Number: | FullyConn 1 Number: 4096

) BE2023/5222

Hi

Step length: (2.2

NOT.

[0048] Those skilled in the art may realize that the units and process steps of each example described in combination with the embodiments disclosed in the present disclosure may be implemented by electronic hardware, computer software, or a combination both. In order to clearly describe the interchangeability of hardware and software, the composition and steps of each example are generally described above according to functions. Whether these functions are performed by hardware or software depends on a specific application and design constraint conditions of the technical solution. Professionals may use different methods to implement the functions described for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0049] In the several embodiments provided in the application, it is understood that the systems, devices and methods disclosed can be realized in other modes. For example, the embodiment of the apparatus described above is only given by way of example. For example, the division of units is only a logical function division, other modes of division can be adopted in practice.

For example, a plurality of units or components may be combined or integrated into another system, or certain features may be omitted or not performed. In addition, the mutual coupling or direct coupling or communication connection displayed or exposed may be an indirect coupling or communication connection of devices or units through certain interfaces, and may

0 BE2023/5222 also be in electrical, mechanical or other forms.

[0050] Units shown as separate components may or may not be physically separate, and components shown as units may or may not be physical units, i.e., components may be positioned in one location or location. can also be distributed across a plurality of network units. The solution objectives of the embodiments of the present invention can be achieved by selecting some or all of the units according to actual needs.

[0051] Additionally, the functional units in the embodiments of the present invention may be integrated into a processing unit, or each of the units may physically exist alone, or two or more units may be integrated into a unit. The integrated unit may be implemented in the form of hardware, or may be implemented in the form of a software functional unit.

[0052] When the integrated unit is produced in the form of a software functional unit and sold or used as an independent product, the integrated unit can be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present invention substantially, or the part of the present invention contributing to the prior art, or part of the technical solution may be incorporated in the form of a software product , and the computer software product is stored in a storage medium, which includes a plurality of instructions enabling the computer equipment (which may be a personal computer, a server, or network equipment and the like) to execute the all or part of the steps in the methods of the embodiments of the present invention.

The aforementioned storage medium includes: various media capable of storing program codes, such as U disk, mobile hard disk, read only memory (ROM), random access memory (RAM), magnetic disk or optical disk .

Claims (9)

5 CLAIMS 1. Method for identifying fish in fish farming, in which the method comprises: developing a bank of fish image samples; extracting shapes and positions of fish heads and fish tails from fish image samples to construct a joint fish identification feature bank of fish bodies, fish heads and fish tails; and developing a fish identification transfer learning model, pre-training the fish image sample bank according to the fish identification feature bank, and identifying fish in fish farming through transfer learning model of fish identification. 2. Method for identifying fish in fish farming according to claim 1, further comprising constructing a small data set of samples of samples to be tested; training the pre-learned fish identification transfer model again using the small sample set; and inputting the resulting fish images into the fish identification transfer model to identify individual fish and fish species. 3. Method for identifying fish in fish farming according to claim 2, further comprising developing a minimum bounding rectangle of individual fish, and then expanding the horizontal axis and the vertical axis of the rectangle minimum delimitation of n pixels respectively to develop a first identification zone; identifying positions of fish heads and fish tails in the first identification zone; And 1 BE2023/5222 setting the direction of a line connecting fish tails to fish heads as the direction of fish body shape by taking the center point of the minimum bounding rectangle of fish heads and fish tails fish as a directional positioning point. 4. Method for identifying fish in fish farming according to claim 2, wherein sets of training and test samples are constructed based on target features of fish body, fish head and fish tail , and images in the sample sets are mirror image calibrated by rotating -30°, -60°, 30° and 60° and scaling 1:2, 1:0.75, 1: 0.5 and 1:0.25 respectively to increase the number of samples. 5. A fish identification method in fish farming according to claim 1, wherein a fully connected layer is added at the end of a VGG19 pre-trained model to develop the fish identification transfer learning model. 6. Method for identifying fish in fish farming according to claim 1, in which the fish are classified into cyprinoideae, hypophthalmichthyinae, cobitinae, acipenserinae and bagridae to develop the bank of fish samples of families and genera. 7. Fish identification platform in fish farming, wherein the platform includes a server, the server having a memory and a processor coupled to the memory, the processor configured to execute instructions stored in the memory, the processor executing the following operations : development of a bank of fish image samples; extracting shapes and positions of fish heads and fish tails from sample fish images to construct a joint fish identification feature bank of fish bodies, fish heads and tails 5 of fish; developing a fish identification transfer learning model, pre-training the fish image sample bank according to the fish identification feature bank, and identifying fish in fish farming through fish identification transfer learning model; constructing a small dataset of sample samples; training the pre-learned fish identification transfer model again using the small sample set; and inputting the obtained fish images into the fish identification transfer model to identify individual fish. 8. Fish farming fish identification terminal, wherein the terminal has an image acquisition device for acquiring fish farming images, and the terminal uploads the acquired fish images to the fish identification platform in fish farming according to claim 7, and the platform server returns the results of the identification to the identification terminal after identification of the fish. 9. Storage medium storing a computer program thereon, wherein when the program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.