Method, system, storage medium and device for identifying aquatic animals in aquaculture water body

Abstract

The application discloses a kind of aquaculture water body in aquatic animal identification method, system, storage medium and equipment, it is related to image recognition technical field, the method is by camera module deployed in water body interior and is gathered multi-view image construction image set, and by floating type communication device synchronization space-time metadata stable transmission to cloud end;In cloud end, sequentially execute the unsupervised screening based on image content features, color self-adaptive correction according to the optical characteristics of water body and the frequency domain enhancement processing combined with semantic guidance, form three-level preprocessing pipeline, significantly improve image quality;Finally, the optimized image is input into pre-trained underwater target detection model, and the high-precision aquatic animal identification result is output.The method effectively solves the problems of poor data transmission stability, low target feature recognition and insufficient real-time identification in complex water environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, in particular to a method and system for identifying aquatic animals in aquaculture water, a storage medium and equipment. BACKGROUND

[0002] With the large-scale development of global aquaculture industry, accurate monitoring of the growth status of cultured organisms has become a core requirement for industrial upgrading. Traditional methods relying on manual inspection have low efficiency, limited coverage, and large subjective errors, which cannot meet the real-time monitoring requirements of large-scale high-density aquaculture scenarios. Although intelligent recognition technology based on computer vision has been widely used in terrestrial scenarios, it has long faced challenges in adapting to the technical requirements of aquatic animal recognition due to the complex optical properties of water environments (such as light scattering, low contrast, and turbid medium interference) and the physical limitations of underwater device deployment.

[0003] Existing technologies in the field of image processing often use global histogram equalization or traditional filtering algorithms, which cannot adapt to the dynamic turbidity changes of water bodies, easily cause local overexposure, noise amplification, or boundary artifacts, and lack automated invalid image screening mechanisms, resulting in a large amount of redundant calculations. In addition, the recognition model directly applies a general convolutional neural network (CNN), which is sensitive to underwater light scattering and low-contrast feature degradation, leading to low recognition accuracy and high miss rate, making it difficult to support real-time decision analysis. Overall, there are defects such as poor transmission stability, low target recognition rate, and insufficient recognition efficiency. SUMMARY

[0004] To overcome the deficiencies of the prior art, the present application aims to provide a method and system for identifying aquatic animals in aquaculture water, a storage medium and equipment, which aims to solve the above problems described in the prior art.

[0005] The first aspect of the present application provides a method for identifying aquatic animals in aquaculture water, the method comprising:

[0006] By deploying a camera module inside the aquaculture water, image data of multiple viewing angles in the current state of the aquaculture water is collected, and an image set is constructed based on the image data;

[0007] The communication device floating on the aquaculture water receives the image set in real time, transmits the image set to the cloud through the built-in communication unit, and matches the spatio-temporal metadata corresponding to each image in the image set in real time through the built-in positioning system to synchronize transmission to the cloud;

[0008] In the cloud, the received image set is preprocessed, including at least unsupervised screening based on image content features, image color adaptive correction based on water optical properties, and frequency domain enhancement processing for underwater noise and target features, to generate a labeled image set;

[0009] inputting the labeled image set into a pre-trained underwater target detection model for recognition analysis, and outputting an aquatic animal recognition result.

[0010] According to an aspect of the above technical solution, the step of unsupervised screening based on image content features comprises:

[0011] calculating the brightness variance and texture entropy of each image in the image set, and removing invalid images in the image set based on preset brightness variance threshold and texture entropy threshold to obtain a preliminary screening image set;

[0012] performing unsupervised classification on the preliminary screening image set using a clustering algorithm, filtering out single background category images in the classification result based on the brightness variance and edge density features of the images, and obtaining an effective target image set.

[0013] According to an aspect of the above technical solution, the step of performing image color adaptive correction according to the optical properties of the water body comprises:

[0014] analyzing the global color statistical features and texture patterns of the images according to the effective target image set, and outputting a probability distribution representing the type of the water body;

[0015] determining the dominant type of the water body according to the probability distribution of the type of the water body, and selecting a corresponding color correction algorithm accordingly;

[0016] based on the contrast index of the corrected image, dynamically adjusting the gain parameter of the color correction algorithm through a feedback loop to optimize the correction effect;

[0017] If the aquaculture water body is a seawater type, a blue light compensation algorithm based on a light attenuation physical model is activated, and if the aquaculture water body is a freshwater type, a white balance algorithm is enabled, and the image is enhanced in the green spectrum in the HSV color space.

[0018] According to an aspect of the above technical solution, the step of performing frequency domain enhancement processing on the underwater noise and target features comprises:

[0019] performing discrete wavelet transform on the image after color adaptive correction to obtain low-frequency multi-scale frequency domain components, including background information components and high-frequency detail feature components;

[0020] performing edge and texture enhancement on the high-frequency detail feature components, and performing adaptive threshold processing based on filtering on the low-frequency background information components to suppress background noise;

[0021] The pre-trained semantic segmentation model is used to generate an image mask corresponding to a target region in the aquaculture water body, the image mask is used for semantic filtering on the processed multi-scale frequency domain components, non-target region frequency components are filtered out, and the filtered multi-scale frequency domain components are reconstructed into an enhanced spatial domain image through inverse discrete wavelet transform.

[0022] According to an aspect of the above technical solution, the step of inputting the labeled image set into the pre-trained underwater target detection model for identification analysis and outputting the aquatic animal identification result includes:

[0023] The labeled image set containing the labeled images is imported into the pre-trained underwater target detection model for multi-round training, the image configuration parameters are modified, and a weight file is obtained.

[0024] According to the weight file, the labeled images in the labeled image set are subjected to image recognition, and the corresponding aquatic animal identification result is output.

[0025] The aquatic animal identification result at least includes the species category of the aquatic animals and the number of individuals corresponding to each species category.

[0026] According to an aspect of the above technical solution, the step of inputting the labeled image set into the pre-trained underwater target detection model for identification analysis and outputting the aquatic animal identification result further includes:

[0027] The aquatic animal identification result is correspondingly bound to the spatiotemporal metadata of the corresponding labeled images in the labeled image set, and the identification result in a preset format is output and stored in a distributed database.

[0028] The spatiotemporal metadata includes a timestamp and location information corresponding to the image acquisition process.

[0029] According to an aspect of the above technical solution, the method further includes:

[0030] The communication strength between the communication device floating on the aquaculture water body and the cloud is detected in real time, and a data transmission mode is selected based on the communication strength.

[0031] When the communication strength between the communication device floating on the aquaculture water body and the cloud is less than a preset communication strength threshold, the image set and the spatiotemporal metadata are compressed through the communication device, and the compressed data is transmitted in a short message transmission mode.

[0032] When the cloud receives the compressed data, the compressed data is reorganized to obtain the image set and the spatiotemporal metadata.

[0033] The second aspect of the present application provides an aquatic animal identification system in a culture water body, which is applied to the method in the above technical solution, and comprises:

[0034] A data acquisition module is configured to acquire image data of multiple perspectives in the current state of the culture water body through a camera module deployed in the culture water body, and to construct an image set according to the image data;

[0035] A data transmission module is configured to receive the image set in real time through a communication device floating on the culture water body, to transmit the image set to the cloud through a built-in communication unit, and to match time-space metadata corresponding to each image in the image set in real time through a built-in positioning system and transmit the time-space metadata to the cloud in synchronization;

[0036] A data processing module is configured to preprocess the received image set in the cloud, including at least unsupervised filtering based on image content features, image color self-adaptive correction according to water optical properties, and frequency domain enhancement processing for underwater noise and target features, to generate a labeled image set;

[0037] An identification execution module is configured to input the labeled image set into a pre-trained underwater target detection model for identification analysis, and to output an aquatic animal identification result.

[0038] The third aspect of the present application provides a readable storage medium having computer instructions stored thereon, which are executed by a processor to implement the steps of the method in the above technical solution.

[0039] The fourth aspect of the present application provides an electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the method in the above technical solution when executing the program.

[0040] Compared with the prior art, the aquatic animal identification method, system, storage medium and device of the present application have the following beneficial effects:

[0041] The method shown in the present application collects image data of multiple views in the current state of the aquaculture water body through the camera module deployed in the aquaculture water body, and constructs an image set according to the image data; the communication device floating on the aquaculture water body receives the image set in real time, transmits the image set to the cloud through the built-in communication unit, and matches the space-time metadata corresponding to each image in the image set in real time through the built-in positioning system to synchronize transmission to the cloud; in the cloud, the unsupervised image classification method is used to filter invalid images in the image set, the remaining images are preprocessed, and an image set containing labeled images is output; the image set containing labeled images is imported into the pre-trained deep learning model for image recognition, and the corresponding aquatic animal recognition result is output. The method shown in the present application solves the problems of low data transmission stability, low target feature distinguishability and low real-time high-precision recognition in complex water environment. BRIEF DESCRIPTION OF DRAWINGS

[0042] The above and / or additional aspects and advantages of the present application will become apparent and more readily appreciated from the following description of the embodiments, taken in conjunction with the accompanying drawings, in which:

[0043] Figure 1 The flowchart of the aquatic animal recognition method in the aquaculture water body in an embodiment of the present application is shown.

[0044] Figure 2 The structural block diagram of the aquatic animal recognition system in the aquaculture water body in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0045] In order to make the purpose, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below in conjunction with the drawings. However, the present application can be realized in many different forms, and is not limited to the embodiments described herein. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present application more thorough and comprehensive.

[0046] It should be noted that when an element is referred to as being "fixed to" another element, it can be directly on the other element or there can be a middle element. When an element is referred to as being "connected" to another element, it can be directly connected to the other element or there can be a middle element. The terms "vertical", "horizontal", "left", "right" and the like used herein are for illustrative purposes only.

[0047] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in the description herein is for describing particular embodiments only and is not intended to be limiting of the application. The use herein of the terms "and / or" includes a set of one or more associated listed items.

[0048] Embodiment one

[0049] Please refer to Figure 1 The first embodiment of the present application provides a method for identifying aquatic animals in a culture water body, which comprises steps S10-S40:

[0050] Step S10, through the camera module arranged inside the culture water body, image data of multiple views under the current state of the culture water body is collected, and an image set is constructed according to the image data.

[0051] First of all, it needs to be pointed out that this method is suitable for high-precision identification of aquatic animals in a culture water body, such as identifying the species, distribution, and individual quantity of aquatic animals under the current state of the culture water body, wherein the culture water body includes but is not limited to ponds, reservoirs, and paddy fields.

[0052] In this embodiment, when identifying aquatic animals in a culture water body, the camera module arranged inside the culture water body will first collect image data of multiple views under the current state of the culture water body, that is, multiple images are collected, and then a corresponding image set is constructed based on this.

[0053] Among them, when image collection is carried out through the camera module, image collection can be carried out based on a preset time interval, for example, the camera module is controlled to carry out image collection according to a preset time interval, of course, the camera module can also be triggered to start image collection based on the detection of the proximity of aquatic animals by an infrared probe.

[0054] Specifically, in this embodiment, the camera module includes multiple underwater array camera modules, each module includes at least four wide-angle cameras, and is arranged in a ring array at the bottom of the culture water body, the overlapping rate of the view angles of adjacent cameras is ≥30%, the working frame rate is set to 0.1fps-0.5fps, the images of aquatic animals are collected according to a preset time interval, such as every 10 minutes, to generate multi-view image data, and then a corresponding image set is constructed based on the multi-view image data.

[0055] Step S20, receiving the image set in real time through the communication device floating on the aquaculture water body, transmitting the image set to the cloud through the built-in communication unit, and matching the space-time metadata corresponding to each image in the image set in real time through the built-in positioning system to transmit to the cloud.

[0056] In this embodiment, after the image set is obtained by the camera module, the image set is received in real time by the communication device, and data transmission is performed by the communication device, and the space-time metadata corresponding to the image data is matched and transmitted in the process.

[0057] In the communication device, the data coupling between Wi-Fi and the host MCU is processed by a Wi-Fi module. After receiving the image set sent by the camera module, the host MCU encapsulates the image data in the image set into an IP data packet and writes a timestamp header. The communication device needs to transmit the above-mentioned IP data packet to the cloud, which is realized based on a double-link physical isolation transmission mechanism, including a 5G channel and a Beidou channel. The 5G channel transmits the IP data packet carrying the timestamp to the cloud through the 5G NR protocol stack based on the built-in communication unit. The Beidou channel separately calls the Beidou RDSS short message function to transmit a simplified positioning packet and perform space-time synchronization calibration. Specifically, the 1PPS second pulse signal output by the Beidou module is connected to the interrupt pin of the host MCU, so that the time deviation between image acquisition time and positioning timestamp is controlled within ±10ms.

[0058] Specifically, the communication device and the cloud have a double-link data transmission channel. The communication device can autonomously select a data transmission mode according to the current communication connection quality. For example, when the communication device and the cloud have good signal strength, specifically good 5G signal strength, the built-in communication unit of the communication device, i.e., the 5G unit, directly transmits the original image set and the corresponding space-time metadata. When the signal strength is general or poor, the image set is compressed and transmitted together with the corresponding space-time metadata.

[0059] Step S30, in the cloud, pre-processing the received image set, including at least unsupervised filtering based on image content features, image color adaptive correction according to water body optical properties, and frequency domain enhancement processing for underwater noise and target features, to generate a labeled image set.

[0060] In the embodiment, the image content features are used for unsupervised screening, the image color is self-adaptively corrected according to the water body optical characteristics, and the frequency domain enhancement processing is performed on the underwater noise and target features, specifically including filtering invalid images in the image set by using an unsupervised image classification method in the cloud, pre-processing the remaining images, and outputting an image set containing labeled images.

[0061] More specifically, it includes:

[0062] The brightness variance and texture entropy of each image in the image set are obtained, and at least part of the invalid images in the image set are removed according to the brightness variance and texture entropy;

[0063] The remaining images in the image set are classified by using a clustering algorithm to remove at least part of the single-species category images again;

[0064] The remaining images in the image set are divided into grids, the contrast of each grid sub-block is independently limited for adaptive histogram equalization, and the boundary artifacts between adjacent grid sub-blocks are eliminated by using bilinear interpolation to obtain intermediate images;

[0065] The intermediate images are subjected to image enhancement processing according to a preset underwater convolutional neural network, and an image set containing labeled images is output.

[0066] Specifically, in the embodiment, in the cloud, first, the unsupervised image processing unit in the preprocessing module is used to quickly remove obvious invalid samples in the image set, such as images obviously not containing aquatic animals, by using simple indicators such as brightness variance and texture entropy, and the K-means clustering algorithm is used to unsupervisedly classify the remaining images in the image set to further remove background single-species category images. Secondly, based on the image denoising unit in the preprocessing module, the CLAHE algorithm based on 8x8 grid is implemented, each grid sub-block independently performs contrast limited adaptive histogram equalization, and then the sub-blocks are fused by using bilinear interpolation to eliminate the boundary artifacts between the blocks, so as to enhance the contrast of the images. Finally, based on the image enhancement unit in the preprocessing module, an underwater convolutional neural network (UWCNN) is used to eliminate the turbidity effect of the images to enhance the image quality, which specifically includes:

[0067] The dilation rate of the dilation convolution layer is [1, 2, 4, 8, 16], which is used to expand the receptive field and capture multi-scale context features; and a triple residual connection structure is added at the tail, which is used to suppress gradient disappearance and accelerate network convergence, and outputs high-quality labeled images.

[0068] Step S40, inputting the labeled image set into a pre-trained underwater target detection model for recognition analysis, and outputting aquatic animal recognition results.

[0069] It should be noted that after the unsupervised image classification method is used for preprocessing in the cloud to obtain the labeled images, formal image recognition, i.e., aquatic animal recognition, is performed based on the labeled images, and then corresponding aquatic animal recognition results, such as species and quantity, are output.

[0070] In the embodiment, the step of inputting the labeled image set into the pre-trained underwater target detection model for recognition analysis and outputting the aquatic animal recognition result includes:

[0071] The labeled image set containing the labeled images is imported into the pre-trained underwater target detection model for multi-round training, and the image configuration parameters are modified to obtain a weight file.

[0072] According to the weight file, the labeled images in the labeled image set are subjected to image recognition, and corresponding aquatic animal recognition results are output; wherein the aquatic animal recognition results at least include species categories of aquatic animals and individual quantities corresponding to each species category.

[0073] In addition, the step of inputting the labeled image set into the pre-trained underwater target detection model for recognition analysis and outputting the aquatic animal recognition result further includes:

[0074] The aquatic animal recognition result is correspondingly bound with the spatiotemporal metadata of the corresponding labeled image in the labeled image set, and a recognition result in a preset format is output and stored in a distributed database; wherein the spatiotemporal metadata includes a timestamp and position information corresponding to the image acquisition process.

[0075] In the embodiment, the underwater target detection model described above is a deep learning model, which is trained by the following steps:

[0076] After obtaining the aquatic animal images, information of the aquatic animal images is labeled by an image labeling tool to obtain image labeling information for each aquatic animal image.

[0077] A sample training set is constructed according to the aquatic animal images and the image labeling information, and the sample training set is divided to obtain a training set, a validation set, and a test set, for example, 7:2:1.

[0078] The aquatic animal images are subjected to a preprocessing operation.

[0079] The aquatic animal images are subjected to a data enhancement operation to expand data samples.

[0080] The data augmentation operation includes at least one or any combination of horizontal flipping, vertical flipping, 90° clockwise flipping, 180° clockwise flipping, 270° clockwise flipping, transposition and image mirroring of the original image data using Python language and OpenCV coding, and in addition, the model generalization capability can be improved by generating training samples with different color deviations and scattering intensities.

[0081] Specifically, in the present embodiment, when identifying aquatic animals, data augmentation is first performed, specifically, the labeled images after labeling are subjected to data augmentation, including generating synthetic data sets of different water types based on an underwater light attenuation model, such as simulating the attenuation ratio of blue-red, blue-green channels, and expanding the labeled data set by rotating and cropping.

[0082] Secondly, model training, the preprocessed data set is input into the YOLOv8 model for multi-round training, based on the VOC2017 model pre-training weight file, by modifying the species category number, data set path and other parameters, finally the accurate weight file is obtained.

[0083] Finally, real-time identification, the image received by the cloud after image noise reduction and image enhancement processing is input into the trained YOLOv8 model for accurate prediction, and the species category, individual quantity and other information of aquatic animals are output.

[0084] More specifically, a large amount of image data taken underwater is processed using an unsupervised image classification method, a double-threshold decision mechanism, i.e., a brightness variance threshold and a texture entropy threshold, is used to filter non-target images, i.e., invalid images without aquatic animals, and a series of classified images are output to facilitate subsequent data processing, and the k-means algorithm is used to perform unsupervised clustering on the remaining images in the image set to preprocess the data, specifically, based on the brightness variance and edge density characteristics, the images are automatically classified into three categories: valid target frames, blank frames and high-blur frames, and the invalid frames, including blank frames and high-blur frames, are automatically removed by setting a clarity threshold.

[0085] And, since the underwater shot image usually has the problem of poor quality, which may affect the recognition effect, image enhancement needs to be performed first. In this embodiment, CLAIHE algorithm is used to improve the image contrast. Specifically, the image is divided into an 8x8 grid, and the CLAHE algorithm is applied to each block independently. Bilinear interpolation is used to eliminate block boundary artifacts. The original image is improved in contrast by the CLAHE algorithm, and the UWCNN algorithm is used to enhance the overall image quality. Specifically, the average value of the RGB three channels of the image is calculated, and then the average value of the three channel averages is taken as the target. The gain coefficient of each channel is calculated respectively, and the pixel value of each channel is adjusted through the gain coefficient, so as to correct the color deviation of the image and provide a more accurate color basis for subsequent image processing.

[0086] During the training process, the collected picture data and the artificially labeled label information are input into the YOLOv8 model, and in the detection stage, the detection result and the true result data information are constantly fitted, and by constantly modifying the hyperparameters, the best weight of the detection model is finally obtained, and the detection process directly outputs the detection result using the weight.

[0087] Among them, the loss function L is as follows:

[0088]

[0089]

[0090] In the formula, , , are weight coefficients for balancing the optimization intensity of different tasks, IoU represents the intersection over union of the predicted box and the real box, p represents the Euclidean distance of the center points of the two boxes, c represents the diagonal length of the minimum closed region covering the two boxes, v represents the width-height ratio penalty term, and a is the weight coefficient thereof, represents the jth real label of the ith sample, represents the probability that the ith sample belongs to the jth class predicted by the model, represents the target real label of the ith predicted box, represents the confidence that the ith predicted box contains the target predicted by the model.

[0091] Among them, in the detection process, any 5 species categories of the species can be input into the network model, and the detection of each picture can be performed through image denoising and image enhancement. In the subsequent specific detection and recognition process, since the weight has been obtained, the weight is input into the YOLOv8 model, and the detection and recognition result can be directly obtained, so that the species category on the picture can be detected.

[0092] Compared with the prior art, the aquaculture water body aquatic animal identification method shown in the embodiment has the beneficial effects that:

[0093] The method shown in the embodiment collects image data of multiple views in the current state in the aquaculture water body through the camera module arranged in the aquaculture water body, and constructs an image set according to the image data; the communication device floating on the aquaculture water body receives the image set in real time, transmits the image set to the cloud through the built-in communication unit, and matches the space-time metadata corresponding to each image in the image set in real time through the built-in positioning system to synchronize transmission to the cloud; in the cloud, the unsupervised image classification method is used to filter invalid images in the image set, the remaining images are preprocessed, and an image set containing labeled images is output; the image set containing labeled images is input into the pre-trained deep learning model for image recognition, and the corresponding aquatic animal identification result is output. The method shown in the embodiment solves the problems of low data transmission stability, low target feature distinguishability and low real-time high-precision identification in complex water body environment.

[0094] Embodiment two

[0095] The second embodiment of the application also provides an aquaculture water body aquatic animal identification method, the method shown in the embodiment is basically similar to the method shown in the first embodiment, and the difference is that:

[0096] In the embodiment, the step of unsupervised screening based on image content features includes:

[0097] The brightness variance and texture entropy of each image in the image set are calculated, and based on the preset brightness variance threshold and texture entropy threshold, invalid images in the image set are removed to obtain a preliminary screening image set;

[0098] The clustering algorithm is used for unsupervised classification of the preliminary screening image set, and based on the brightness variance and edge density features of the image, the single background category images in the classification result are filtered out to obtain an effective target image set.

[0099] The step of performing image color adaptive correction according to the optical properties of the water body includes:

[0100] According to the effective target image set, the global color statistical features and texture patterns of the image are analyzed, and the probability distribution representing the water body type is output;

[0101] According to the probability distribution of the water body type, the dominant water body type of the water body is determined, and the corresponding color correction algorithm is selected;

[0102] Based on the contrast index of the corrected image, the gain parameter of the color correction algorithm is dynamically adjusted through a feedback loop to optimize the correction effect;

[0103] wherein, if the aquaculture water body is of seawater type, a blue light compensation algorithm based on a light attenuation physical model is activated, and if the aquaculture water body is of freshwater type, a white balance algorithm is enabled, and the image is subjected to green spectrum enhancement in the HSV color space.

[0104] In addition, the step of performing frequency domain enhancement processing on the underwater noise and target features comprises:

[0105] Discrete wavelet transform is performed on the image after color self-adaptive correction, and low-frequency multi-scale frequency domain components, including background information components and high-frequency detail feature components, are obtained by decomposition.

[0106] The high-frequency detail feature components are subjected to edge and texture enhancement, and the low-frequency background information components are subjected to adaptive threshold processing based on filtering to suppress background noise.

[0107] A pre-trained semantic segmentation model is used to generate an image mask corresponding to the target region in the aquaculture water body, the processed multi-scale frequency domain components are subjected to semantic filtering based on the image mask to filter out frequency components of non-target regions, and the filtered multi-scale frequency domain components are reconstructed into enhanced spatial domain images through inverse discrete wavelet transform.

[0108] Specifically, the core of the method shown in the embodiment is consistent with that of the first embodiment, but its core contribution lies in the deep optimization of the cloud preprocessing link, which significantly improves the quality of underwater images through the intelligent preprocessing pipeline and precise connection of the preprocessing process, laying a solid foundation for subsequent accurate recognition.

[0109] First, in terms of unsupervised screening, a two-stage filtering strategy from coarse to fine is adopted. The first stage of filtering includes two quickly computable quantitative indicators based on image brightness variance and texture entropy, which efficiently batch-rejects invalid images with overexposure, overdarkness or blurred content, significantly reducing the computational burden from the source. Subsequently, the second stage of filtering uses a clustering algorithm to unsupervisedly classify the images that pass the preliminary screening, intelligently identifies and filters out images that have single background and do not contain valid targets based on features such as brightness variance and edge density. The above two stages of screening together ensure that only content-related and quality-considerable images can enter the subsequent processing link.

[0110] ​Secondly, the most critical color adaptive correction first intelligently analyzes the color and texture features of the image through a lightweight convolutional network, judges the type of water body, such as seawater or freshwater, and then differentially corrects according to different physical optical models. Among them, for seawater environment, the blue light compensation algorithm based on light attenuation physical model is activated to compensate the loss of red light band; and for freshwater environment, the white balance algorithm is enabled, and further specific enhancement of green spectrum in HSV color space is performed. More specifically, by introducing a feedback loop mechanism, the parameters can also be dynamically fine-tuned according to the actual effect after correction, so as to adapt to the dynamic changes of the water environment and ensure the continuous accuracy of color restoration.

[0111] Finally, the frequency domain enhancement can effectively improve the image processing to a new dimension. The corrected image is decomposed into frequency domain components of different scales through discrete wavelet transform. On this basis, the high-frequency components carrying detailed information are enhanced to highlight the target outline, while the low-frequency components carrying background information are subjected to noise suppression. More importantly, by introducing a semantic guided filtering mechanism, a pre-trained semantic segmentation model is used to identify potential aquatic animal regions and generate accurate image masks, so as to selectively enhance the target area in the frequency domain, filter out the frequency components of interference factors such as bubbles and aquatic plants, and finally reconstruct the spatial domain image with significantly improved clarity through inverse transform.

[0112] In summary, the method shown in the embodiment can effectively solve the problem of low underwater aquatic animal recognition accuracy caused by poor underwater image quality, color distortion, and large noise interference in the prior art through the above-mentioned preprocessing cooperation.

[0113] Embodiment Three

[0114] Please refer to Figure 2 The third embodiment of the present application provides a cultured water body aquatic animal recognition system, which is applied to the method of any one of the above-mentioned embodiments. The system comprises:

[0115] A data acquisition module 10 is configured to acquire image data of multiple views in the current state of the cultured water body through a camera module deployed inside the cultured water body, and construct an image set according to the image data;

[0116] A data transmission module 20 is configured to receive the image set in real time through a communication device floating on the cultured water body, transmit the image set to the cloud through an internal communication unit, and match the spatio-temporal metadata corresponding to each image in the image set in real time through an internal positioning system to synchronize transmission to the cloud;

[0117] The data processing module 30 is configured to perform preprocessing on the received image set in the cloud, including at least unsupervised filtering based on image content features, image color self-adaptive correction according to water body optical characteristics, and frequency domain enhancement processing for underwater noise and target features, to generate a labeled image set.

[0118] The recognition execution module 40 is configured to input the labeled image set into a pre-trained underwater target detection model for recognition analysis, and output a water animal recognition result.

[0119] Compared with the prior art, the water animal recognition system in the embodiment has the following beneficial effects:

[0120] The system in the embodiment collects image data of multiple perspectives in the current state of the aquaculture water body through the camera module arranged in the aquaculture water body, and constructs an image set according to the image data. The communication device floating on the aquaculture water body receives the image set in real time, the built-in communication unit transmits the image set to the cloud, and the built-in positioning system matches the space-time metadata corresponding to each image in the image set in real time to synchronize transmission to the cloud. In the cloud, the unsupervised image classification method is used to filter invalid images in the image set, the remaining images are preprocessed, and an image set containing labeled images is output. The image set containing labeled images is imported into a pre-trained deep learning model for image recognition, and a corresponding water animal recognition result is output. The system in the embodiment solves the problems of low data transmission stability, low target feature distinguishability, and low real-time high-precision recognition in a complex water environment.

[0121] Embodiment Four

[0122] The fourth embodiment of the present application provides a readable storage medium having computer instructions stored thereon, the instructions being executed by a processor to implement the steps of the method described in any of the above embodiments.

[0123] Embodiment Five

[0124] The fifth embodiment of the present application provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described in any of the above embodiments.

[0125] In the description of the present specification, the description of the terms "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0126] The above-described embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but cannot be understood as a limitation on the patent scope of the present application. It should be noted that, for those of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. An aquaculture water body method of identifying aquatic animals, characterized by, The method includes: By deploying camera modules inside the aquaculture water body, image data from multiple perspectives under the current state of the aquaculture water body is collected, and an image set is constructed based on the image data; The image set is received in real time by a communication device floating on the aquaculture water, the image set is transmitted to the cloud through a built-in communication unit, and the spatiotemporal metadata corresponding to each image in the image set is matched in real time through a built-in positioning system and transmitted synchronously to the cloud. In the cloud, the received image set is preprocessed, including at least unsupervised screening based on image content features, adaptive color correction of images based on the optical properties of water bodies, and frequency domain enhancement processing for underwater noise and target features, in order to generate an annotated image set. The labeled image set is input into a pre-trained underwater target detection model for recognition and analysis, and the aquatic animal recognition results are output. The unsupervised screening steps based on image content features include: Calculate the brightness variance and texture entropy of each image in the image set, and based on preset brightness variance thresholds and texture entropy thresholds, remove invalid images from the image set to obtain a preliminary filtered image set. Clustering algorithms are used to perform unsupervised classification on the initially screened image set. Based on the brightness variance and edge density features of the images, images with a single background category in the classification results are filtered out to obtain an effective target image set. The step of adaptive color correction of the image based on the optical properties of the water body includes: Based on the effective target image set, analyze the global color statistical features and texture patterns of the images, and output the probability distribution representing the water body type; Based on the probability distribution of the water body types, the dominant water body type is determined, and the corresponding color correction algorithm is selected accordingly. Based on the contrast index of the corrected image, the gain parameter of the color correction algorithm is dynamically adjusted through a feedback loop to optimize the correction effect. If the aquaculture water is seawater, the blue light compensation algorithm based on the physical model of light attenuation is activated; if the aquaculture water is freshwater, the white balance algorithm is enabled, and the image is enhanced with green spectrum in the HSV color space. The steps for frequency domain enhancement processing of underwater noise and target features include: Discrete wavelet transform is performed on the color-adaptive corrected image to decompose it into multi-scale frequency domain components, including low-frequency background information components and high-frequency detail feature components. Edge and texture enhancement is performed on the high-frequency detail feature components, while adaptive thresholding based on filtering is applied to the low-frequency background information components to suppress background noise. An image mask corresponding to the target region in the aquaculture water body is generated using a pre-trained semantic segmentation model. Based on the image mask, semantic filtering is performed on the processed multi-scale frequency domain components to filter out frequency components of non-target regions. The filtered multi-scale frequency domain components are then reconstructed into an enhanced spatial domain image through inverse discrete wavelet transform.

2. The method of claim 1, wherein The steps of inputting the labeled image set into a pre-trained underwater target detection model for recognition and analysis, and outputting aquatic animal recognition results, include: The labeled image set containing the labeled images is imported into a pre-trained underwater target detection model for multi-round training, and the image configuration parameters are modified to obtain a weight file; According to the weight file, the labeled images in the labeled image set are subjected to image recognition, and the corresponding aquatic animal recognition result is outputted; The aquatic animal recognition result at least includes the species category of aquatic animals and the individual quantity corresponding to each species category.

3. The method of claim 2, wherein The step of inputting the labeled image set into the pre-trained underwater target detection model for recognition analysis and outputting the aquatic animal recognition result further includes: The aquatic animal recognition result is correspondingly bound with the spatio-temporal metadata of the corresponding labeled image in the labeled image set, and the recognition result in a preset format is outputted and stored into a distributed database; The spatio-temporal metadata includes the timestamp and location information corresponding to the image acquisition process.

4. The method of claim 1-3, wherein The method further includes: The communication strength between the communication device floating on the aquaculture water body and the cloud is detected in real time, and the data transmission mode is selected based on the communication strength; When the communication strength between the communication device floating on the aquaculture water body and the cloud is less than a preset communication strength threshold, the image set and the spatio-temporal metadata are compressed by the communication device, and the compressed data is transmitted by using the short message transmission mode; When the cloud receives the compressed data, the compressed data is reorganized to obtain the image set and the spatio-temporal metadata.

5. An aquaculture water body water animal identification system, characterized by, The system is applied to the method of any one of claims 1-4, and the system comprises: A data acquisition module is configured to acquire image data of multiple perspectives in the current state of the aquaculture water body by a camera module deployed inside the aquaculture water body, and construct an image set according to the image data; A data transmission module is configured to receive the image set in real time by a communication device floating on the aquaculture water body, transmit the image set to the cloud by a built-in communication unit, and match the spatio-temporal metadata corresponding to each image in the image set in real time by a built-in positioning system to synchronously transmit to the cloud; A data processing module is configured to preprocess the received image set in the cloud, including at least unsupervised screening based on image content features, image color self-adaptive correction according to water optical properties, and frequency domain enhancement processing for underwater noise and target features, to generate a labeled image set; An identification execution module is configured to input the labeled image set into a pre-trained underwater target detection model for recognition analysis, and output an aquatic animal recognition result.

6. A readable storage medium, having stored thereon computer instructions, characterized in that, The instruction is executed by the processor to implement the steps of the method of any one of claims 1-4.

7. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method of any one of claims 1-4.

Images