High-definition acquisition and identification method for fish target in turbid water body based on polarization camera shooting and deep learning

Abstract

The invention discloses a high-definition acquisition and identification method for a fish target in a turbid water body based on polarization camera shooting and deep learning. The method comprises the following steps: S1, obtaining an original image with multiple polarization states; s2, carrying out comprehensive image quality scoring; s3, judging whether the image quality score is greater than a preset threshold value or not, and executing different recognition paths; s4, outputting a final recognition result according to the selected path; and S5, storing the identification result and the optimization parameters in a database. Compared with the prior art, the high-definition acquisition and identification method for the fish target in the turbid water body based on polarization camera shooting and deep learning has the advantages that the method is used for the turbid water body, has double-path identification and closed-loop optimization, and realizes high-definition acquisition and reliable identification of the fish target.

Description

A High-Resolution Acquisition and Recognition Method for Fish Targets in Turbid Water Based on Polarization Imaging and Deep Learning Technical Field

[0001] This invention relates to the field of underwater target recognition technology, specifically to a method for high-definition acquisition and recognition of fish targets in turbid waters based on polarization imaging and deep learning. Background Technology

[0002] In natural or artificial water bodies, the presence of suspended particles, plankton, and organic matter often results in highly turbid water. Suspended particles significantly scatter and absorb light, leading to blurry images, low contrast, and difficulty in identifying target features in images captured by traditional underwater camera equipment. The accuracy rate for fish identification is generally below 60%.

[0003] In existing technologies, some studies have attempted to suppress scattered light using polarization imaging. For example, linearly polarized illumination combined with an analyzer is used to enhance the target contour through Stokes parametric calculations. However, single-polarization images still struggle to recover detailed textures, and under extremely low visibility conditions, even after polarization processing, the image information is still insufficient to support reliable identification.

[0004] Therefore, there is an urgent need for a method for acquiring and identifying fish targets that can adapt to turbid water environments, balance image quality and recognition efficiency, and have continuous optimization capabilities. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to overcome the above-mentioned technical defects and provide a high-definition acquisition and recognition method for fish targets in turbid water bodies based on polarization imaging and deep learning, which features dual-path recognition and closed-loop optimization, and achieves high-definition acquisition and reliable recognition of fish targets.

[0006] To address the aforementioned technical problems, the present invention provides the following technical solution: a method for high-resolution acquisition and recognition of fish targets in turbid water based on polarization imaging and deep learning, comprising the following steps: S1: Based on an imaging device equipped with an adjustable polarizer, images of the target area in turbid water are simultaneously acquired at at least three different polarization angles to obtain original images with multiple polarization states; S2: The original images acquired in S1 are processed to generate preliminary enhanced images, and a comprehensive image quality score is performed on the preliminary enhanced images; S3: It is determined whether the image quality score is greater than a preset threshold; when the image quality score is greater than the preset threshold, a first recognition path is executed, and the location, category, and high-resolution reconstructed image of the fish target are output; when the image quality score is less than the preset threshold, a second recognition path is executed, and a high-confidence AI inference engine is used to generate a fish category prediction result with high confidence; S4: According to the selected path, the final recognition result is output; S5: According to the final recognition result, the polarization angle parameters of the imaging device, the processing strategy of the preliminary enhanced image, or the confidence threshold of the high-confidence AI inference engine are optimized through closed-loop feedback, and the recognition result and optimization parameters are stored in a database.

[0007] Preferably, the adjustable polarizer in S1 is an electrically controlled rotating polarizer or a liquid crystal adjustable polarizer, and its polarization angle adjustment accuracy is ≤1°.

[0008] Preferably, the processing of the original image in S2 includes preprocessing and preliminary enhancement; the preprocessing includes dark current correction, non-uniformity correction, geometric distortion correction and white balance adjustment of the original image with multiple polarization states; the preliminary enhancement includes extracting polarization features and fusing them to generate an enhanced image.

[0009] Preferably, the overall quality score in S2 is calculated based on a weighted average of sharpness, contrast, and polarization feature significance indicators.

[0010] Preferably, the first recognition path in S3 includes inputting the initially enhanced image into a high-definition reconstruction network to generate a high-definition image, and inputting the high-definition image into a multi-scale attention fusion target detection network to identify the location and category of fish.

[0011] Preferably, in S3, the high-confidence AI inference engine acquires: real-time environmental sensing data, statistics on the frequency of fish occurrence and species distribution in the same geographical area within similar time windows in historical observation databases, aquatic organism knowledge graphs, and weak image features; after calculating the posterior probability of each candidate fish category, the category with the highest probability and exceeding the confidence threshold is selected as the prediction result.

[0012] Preferably, the weak image features include the rough outline of the target region, direction and speed of motion, polarization contrast response intensity, color distribution trend and size range extracted from the preliminary enhanced image; the weak image features are extracted by a lightweight convolutional neural network and encoded as structured vectors input to a high-confidence AI inference engine.

[0013] Preferably, the polarization features initially enhanced in S2 include intensity, degree of polarization, and polarization angle, and the three are fused to generate an initial enhanced image that highlights the difference between the target and the background.

[0014] Preferably, the closed-loop feedback optimization in S5 includes: adjusting the feature fusion weights of the high-definition reconstruction network when the bounding box localization error of the first path recognition increases; and updating the historical database weights of the high-confidence AI inference engine when the confidence of the second path prediction decreases.

[0015] The advantages of this invention compared to existing technologies are as follows: This invention utilizes an adjustable polarizer to simultaneously acquire images at at least three angles, effectively separating target reflected light from background scattered light, significantly improving the contrast and feature recognition of turbid water images; This invention employs a dual-path adaptive recognition strategy, where high-quality images are accurately located and classified through high-definition reconstruction and a multi-scale attention network, while low-quality images are reliably predicted by fusing multi-source data through a high-confidence AI inference engine; This invention provides a highly efficient end-to-end solution for aquaculture, ecological monitoring, and other scenarios, from image acquisition to intelligent recognition, significantly improving the accuracy and stability of fish target acquisition and recognition in turbid water. Attached Figure Description

[0016] Figure 1 is a flowchart illustrating a method for high-resolution acquisition and identification of fish targets in turbid water based on polarization photography and deep learning. Detailed Implementation

[0017] The present invention will now be described in further detail with reference to the accompanying drawings.

[0018] Referring to Figure 1, a method for high-resolution acquisition and recognition of fish targets in turbid water based on polarization imaging and deep learning includes the following steps: S1: Based on an imaging device equipped with an adjustable polarizer, images of the target area in the turbid water are simultaneously acquired at at least three different polarization angles to obtain original images with multiple polarization states; S2: The original images acquired in S1 are processed to generate preliminary enhanced images, and a comprehensive image quality score is performed on the preliminary enhanced images; S3: It is determined whether the image quality score is greater than a preset threshold; when the image quality score is greater than the preset threshold, the first recognition path is executed, and the location, category, and high-resolution reconstructed image of the fish target are output; when the image quality score is less than the preset threshold, the second recognition path is executed, and a high-confidence AI inference engine is used to generate a fish category prediction result with high confidence; S4: According to the selected path, the final recognition result is output; S5: According to the final recognition result, the polarization angle parameters of the imaging device, the processing strategy of the preliminary enhanced image, or the confidence threshold of the high-confidence AI inference engine are optimized through closed-loop feedback, and the recognition result and optimization parameters are stored in the database.

[0019] In use, the adjustable polarizer in S1 is an electrically controlled rotating polarizer or a liquid crystal adjustable polarizer, with a polarization angle adjustment accuracy of ≤1°. The processing of the original image in S2 includes preprocessing and preliminary enhancement. The preprocessing includes dark current correction, non-uniformity correction, geometric distortion correction, and white balance adjustment for the original image with multiple polarization states. The preliminary enhancement includes extracting polarization features and fusing them to generate an enhanced image. The polarization features extracted in S2 for preliminary enhancement include intensity, degree of polarization, and polarization angle. By fusing these three features, a preliminary enhanced image that highlights the difference between the target and the background is generated.

[0020] In one embodiment: the overall quality score in S2 is calculated based on a weighted average of sharpness, contrast and polarization feature significance.

[0021] In a specific implementation of this invention, the first identification path in S3 includes inputting the initially enhanced image into a high-definition reconstruction network to generate a high-definition image, and inputting the high-definition image into a multi-scale attention fusion target detection network to identify the location and category of fish; the high-confidence AI inference engine in S3 acquires: real-time environmental sensing data, statistics on the frequency of fish occurrence and species distribution in the same geographical area within similar time windows in historical observation databases, aquatic organism knowledge graphs, and weak image features; after calculating the posterior probability of each candidate fish category, the category with the highest probability and exceeding the confidence threshold is selected as the prediction result.

[0022] In practical applications, weak image features include the rough outline of the target region, direction and speed of motion, polarization contrast response intensity, color distribution trend and size range extracted from the preliminary enhanced image; these weak image features are extracted by a lightweight convolutional neural network and encoded as structured vectors input to a high-confidence AI inference engine.

[0023] The closed-loop feedback optimization in S5 includes: adjusting the feature fusion weights of the high-definition reconstruction network when the bounding box localization error of the first path recognition increases; and updating the historical database weights of the high-confidence AI inference engine when the confidence of the second path prediction decreases.

[0024] The contents not described in detail in this specification are existing technologies known to those skilled in the art.

[0025] The working principle of this invention is as follows: Polarization imaging is used to suppress scattered light in turbid water, and multi-polarization images are acquired and fused for enhancement; the overall image quality score is used to determine whether it is sufficient to support visual recognition; if the image quality is high, deep learning is used for high-definition reconstruction and target detection; if the quality is low, a high-confidence AI inference engine is activated, which integrates weak image features, real-time environmental data, historical distribution statistics and aquatic knowledge graphs to calculate the posterior probability of fish categories and output a high-confidence prediction.

[0026] The system provided by this invention also dynamically optimizes polarization parameters, image processing strategies, or inference thresholds based on the recognition effect, forming a closed-loop intelligent mechanism of perception-decision-feedback, realizing dual-path inference, and can output results whether the image is clear or unclear, making it easy to promote and use.

[0027] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0028] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning, characterized in that: The process includes the following steps: S1: Based on an imaging device equipped with an adjustable polarizer, images of the target area in turbid water are simultaneously acquired at at least three different polarization angles to obtain original images with multiple polarization states; S2: The original images acquired in S1 are processed to generate preliminary enhanced images, and the preliminary enhanced images are given a comprehensive image quality score. S3: Determine if the image quality score is greater than a preset threshold; if the image quality score is greater than the preset threshold, execute the first recognition path and output the location, category, and high-resolution reconstructed image of the fish target; if the image quality score is less than the preset threshold, execute the second recognition path and use a high-confidence AI inference engine to generate a fish category prediction result with high confidence. S4: Output the final recognition result based on the selected path; S5: Based on the final recognition result, perform closed-loop feedback optimization on the polarization angle parameter of the imaging device, the processing strategy of the preliminary enhanced image, or the confidence threshold of the high-confidence AI inference engine, and store the recognition result and optimization parameters in the database.

2. The method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 1, characterized in that: The adjustable polarizer in S1 is an electrically controlled rotating polarizer or a liquid crystal adjustable polarizer, and its polarization angle adjustment accuracy is ≤1°.

3. The method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 1, characterized in that: The processing of the original image in S2 includes preprocessing and preliminary enhancement; the preprocessing includes dark current correction, non-uniformity correction, geometric distortion correction and white balance adjustment of the original image with multiple polarization states; the preliminary enhancement includes extracting polarization features and fusing them to generate an enhanced image.

4. The method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 3, characterized in that: The overall quality score in S2 is calculated based on a weighted average of sharpness, contrast, and polarization feature significance.

5. The method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 1, characterized in that: The first identification path in S3 includes inputting the initially enhanced image into a high-definition reconstruction network to generate a high-definition image, and inputting the high-definition image into a multi-scale attention fusion target detection network to identify the location and category of fish.

6. The method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 1, characterized in that: The high-confidence AI inference engine in S3 acquires: real-time environmental sensing data, statistics on the frequency of fish occurrence and species distribution in the same geographical area within similar time windows in historical observation databases, aquatic organism knowledge graphs, and weak image features; After calculating the posterior probability of each candidate fish category, the category with the highest probability that exceeds the confidence threshold is selected as the prediction result.

7. A method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 6, characterized in that: The weak image features include the rough outline of the target region, direction and speed of motion, polarization contrast response intensity, color distribution trend and size range extracted from the preliminary enhanced image; the weak image features are extracted by a lightweight convolutional neural network and encoded as structured vectors input to a high-confidence AI inference engine.

8. A method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 3, characterized in that: The polarization features initially enhanced in S2 include intensity, degree of polarization, and polarization angle. By fusing these three features, a preliminary enhanced image that highlights the difference between the target and the background is generated.

9. A method for high-resolution acquisition and identification of fish targets in turbid water based on polarization imaging and deep learning according to claim 1, characterized in that: The closed-loop feedback optimization in S5 includes: when the bounding box localization error of the first path recognition increases, the feature fusion weights of the high-definition reconstruction network are adjusted; when the confidence of the second path prediction decreases, the historical database weights of the high-confidence AI inference engine are updated.

Images