Intelligent identification method and system for underwater pollutants based on image processing

Abstract

The present application relates to the field of image processing, in particular to an underwater pollutant intelligent identification method and system based on image processing, comprising: collecting multi-region time sequence underwater bottom images through an underwater robot, combining U-Net and YOLO algorithms to obtain suspected garbage positions and probabilities; dividing image windows, calculating anchor point weights based on center point RGB time sequence changes, and constructing turbidity functions using adjacent image window gray scale fluctuations; weightedly fusing window turbidity and anchor point weights to generate regional overall turbidity; dynamically distributing double algorithm weights according to the turbidity to output a fusion probability that the target is garbage; and determining pollution and giving an early warning through a probability threshold. The present application solves the adaptability problem of a single algorithm in turbid / clear water areas, and significantly improves the identification accuracy and anti-interference ability.

Description

Technical Field

[0001] The present invention relates to the field of image processing, and in particular to an image processing-based intelligent identification method and system for underwater pollutants. Background Art

[0002] In today's era of rapid technological advancement, new materials and products are constantly emerging, leading to a surge in solid waste production. According to statistics, millions of tons of solid waste enter waterways worldwide each year, from rivers surrounding cities to the vast oceans. This waste not only releases harmful chemicals into the water, causing eutrophication, disrupting the balance of aquatic ecosystems, and threatening the survival of aquatic life such as fish and corals, but also poses potential risks to shipping safety and water conservancy facilities. Trash in turbid waters has become an "invisible and persistent problem" in water management.

[0003] Existing underwater pollutant detection mainly relies on underwater robots equipped with cameras to collect underwater images and identify them through computer vision algorithms. However, underwater robot recognition cannot adapt to different waters. Commonly used algorithms for underwater robot recognition are U-Net and YOLO. The U-Net algorithm is based on pixel-level analysis and can effectively capture blurred targets through probability output and attention mechanism. It performs well in turbid waters, but may be inefficient due to computational redundancy in clear environments. The YOLO algorithm relies on regular shapes and clear edge features. Its efficient anchor frame mechanism can fully utilize its advantages of fast detection speed and sensitivity to small targets, but it performs poorly in turbid waters because the optical scattering caused by suspended particles in turbid waters significantly reduces the contrast and clarity of the image. Summary of the Invention

[0004] The present invention provides an image processing method and system for underwater pollutant identification to solve the existing problem that the single visual algorithm used by underwater robots for identification cannot adapt to both turbid waters and clear waters, resulting in inaccurate underwater pollutant identification.

[0005] The present invention provides an intelligent underwater pollutant identification method and system based on image processing, which adopts the following technical solutions:

[0006] In a first aspect, the present invention provides an intelligent identification method for underwater pollutants based on image processing, the method comprising the following steps:

[0007] Collect underwater images from different areas and at different times, and classify and archive the images according to the order of the image collection areas and time series;

[0008] According to the identification logic of the U-Net algorithm and the YOLO algorithm for pollutants, the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability of the two algorithms are obtained respectively; set and traverse the image panes of different time series in the same area, and obtain the anchor point change weight of each pane in each image according to the time series change of the pixel values ​​of different color channels at the center point of the pane; calculate the fluctuation of the grayscale value of the pixel points in the same sequence panes of adjacent time series images in the same area, construct the relationship function between turbidity and grayscale value fluctuation, and obtain the turbidity of each pane based on the fluctuation of the grayscale value of the pixel points in different panes; obtain the turbidity of each image according to the turbidity of each pane of each image in each area and the anchor point change weight of each pane of each image in each area, and calculate the average turbidity of all images in the same area as the turbidity of the area; assign the detection probability weight of the suspected garbage of the U-Net algorithm and the YOLO algorithm according to the turbidity of each area, and obtain the probability of detecting the target suspected garbage;

[0009] A probability threshold K is set, and the probability of the detected target being garbage in each area is compared with the probability threshold K to determine whether the detected target is garbage and remind maintenance personnel to take corresponding measures.

[0010] Furthermore, the underwater images of different areas and different time periods are collected, and the images are classified and archived according to the order of the collected image areas and the time sequence, including the specific methods as follows:

[0011] An underwater camera is fixed one meter above the top of the underwater robot. The robot operates in an intermittent motion: after advancing U meters, it stops and captures an image every k seconds while stopped, for a total of T images. After capturing T images, the robot advances U meters again and repeats the process of stopping and capturing until it has captured images of S different areas, each with a size of L x L pixels. Ultimately, all captured images are doubly sorted and traversed, based on the sequence of areas corresponding to the robot's route and the chronological order of image acquisition within the same area.

[0012] Furthermore, the identification logic of the pollutants based on the U-Net algorithm and the YOLO algorithm is used to obtain the detection position of the suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability of the two algorithms, including the specific method:

[0013] For the same image, two algorithms are used to detect the image separately. When the YOLO algorithm successfully detects the target, the position of each YOLO recognition frame is output as the detection position of the suspected garbage. The probability that the target in the recognition frame is garbage is recorded as the detection probability of the YOLO algorithm in the recognition frame. ; Use the U-Net algorithm to calculate the mean probability of all pixels in each YOLO recognition frame being garbage, and use it as the U-Net algorithm detection probability that the target in the corresponding recognition frame of the image is garbage If the YOLO algorithm fails to detect the target, that is, there is no recognition box, the maximum value of the probability that the pixel point in the U-Net full image output is garbage is selected as the U-Net algorithm detection probability that there is garbage in the corresponding area of ​​the image. , and the location of the pixel is used as the detection location of suspected garbage.

[0014] Furthermore, the method of setting and traversing image panes of different time sequences in the same area and obtaining the anchor point change weight of each pane in each image according to the time sequence change of pixel values ​​of different color channels at the center point of the pane includes the following specific methods:

[0015] Divide each L*L image into N*N panes. Starting from the upper left corner of the image, traverse each pane of the image in a snake-like manner to the right. For the nth pane of the tth image in the Sth region, calculate the pixel value change of the center pixel of the pane and the pixel value change of the center pixel of the same sequence pane in the previous image, and obtain the anchor point change parameter of the nth pane of the tth image in the Sth region. , the specific method is as follows:

[0016]

[0017] Where, The R channel pixel value of the center point of the nth pane of the tth image in the sth region, is the size of the G channel of the center point of the nth pane of the tth image in the sth region, is the size of the B channel of the center point of the nth pane of the tth image in the sth region, is the R channel pixel value of the center point of the nth pane of the t-1th image in the sth region, is the size of the G channel of the center point of the nth pane of the t-1th image in the sth region, is the size of the B channel of the center point of the nth pane of the t-1th image in the sth region, represents the anchor point change parameter of the nth pane in the tth image of the sth region;

[0018] Calculate the mean of the anchor point change parameters for all panes of the same sequence in different images of the same region, and record the mean as the anchor point change weight coefficient for all panes of the same sequence in different images of the same region. The specific method is as follows:

[0019]

[0020] represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, represents the anchor point change parameter of the nth pane in the ith image of the sth region, represents the anchor point change weight coefficient of the nth window in the t+1th image of the sth region, where T represents the number of images collected for each region;

[0021] According to the anchor point change weight coefficient of each pane of each image in each region, the anchor point change weight of each pane of each image in each region is obtained as follows:

[0022]

[0023] Where, represents the anchor point change weight of the nth pane of the tth image in the sth region, L represents the side length of the image, N represents the side length of the pane, represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, Represents the anchor point change weight coefficient of the j-th pane in the t-th image of the s-th region.

[0024] Furthermore, the method of calculating the fluctuation of the grayscale values ​​of pixels in the same sequence window panes of adjacent time-series images in the same region, constructing a function relating the turbidity and the grayscale fluctuation, and obtaining the turbidity of each window pane in combination with the fluctuation of the grayscale values ​​of pixels in different window panes includes the following specific methods:

[0025] The average of the absolute values ​​of the differences between the grayscale values ​​of each pixel in the nth window pane of the tth image and the grayscale values ​​of each pixel in the nth window pane of the t-1th image in the same region is recorded as the turbidity parameter of the nth window pane of the tth image in the region. The turbidity parameter of each window pane in the first image is replaced by the average of the turbidity parameters of the window panes in the same sequence of each image in the same region.

[0026] The relationship function between the gray value change and the turbidity degree is established. The specific formula is as follows:

[0027]

[0028] Where g represents the degree of turbidity, represents the hyperparameter, c represents the turbidity parameter, substitute the turbidity parameter of the nth window pane of the tth image in the sth region into the relationship function to obtain the turbidity of the nth window pane of the tth image in the sth region. , Representing the base of the natural logarithm, we further obtain the turbidity level of each pane of each image in each region.

[0029] Furthermore, the method of obtaining the turbidity of each image based on the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region, and calculating the average turbidity of all images in the same region as the turbidity of the region includes the following specific methods:

[0030]

[0031] Where L represents the image side length, N represents the grid side length, represents the turbidity of the t-th image in the s-th region, represents the turbidity of the I-th window pane of the t-th image in the s-th region, represents the anchor point change weight of the I-th pane of the t-th image in the s-th region;

[0032] Calculate the average turbidity of all images in each region as the turbidity of each region. The specific method is as follows:

[0033]

[0034] Where T represents the number of all images in a region, represents the turbidity level of the sth region, Indicates the turbidity level of the J-th image in the s-th region.

[0035] Furthermore, the method of allocating the detection probability weights of the suspected garbage of the U-Net algorithm and the YOLO algorithm according to the turbidity level of each area to obtain the probability of the detected target being suspected garbage includes the following specific methods:

[0036]

[0037] Where, represents the probability of detecting a suspected garbage target in the sth region, represents the turbidity level of the sth region, It represents the probability of the U-Net algorithm detecting the suspected garbage in the sth region. It represents the probability of the YOLO algorithm detecting the suspected garbage when detecting the sth region.

[0038] Furthermore, the probability threshold K is set, and the probability threshold K is compared with the probability of the detected target being garbage in each area to determine whether the detected target is garbage and to remind maintenance personnel to take corresponding measures. The specific method includes:

[0039] Based on the requirements for detection result accuracy in different waters, a probability threshold K is set and compared with the probability P of the detection target in each area being garbage. If P is greater than the probability threshold K, the detection target is determined to be garbage, and maintenance personnel are reminded to salvage and clean it up.

[0040] A second aspect of the present invention provides an underwater pollutant intelligent identification system based on image processing, the system comprising an underwater image acquisition module, a turbidity calculation module, and a pollutant identification module, wherein:

[0041] Underwater image acquisition module, used to collect underwater images of different areas and time periods, and classify and archive the images according to the order of the image areas and time series;

[0042] The turbidity calculation module is used to obtain the detection position of suspected garbage in the same image and the U-Net detection probability and YOLO detection probability of the two algorithms based on the recognition logic of pollutants of the U-Net algorithm and the YOLO algorithm respectively; set and traverse the image panes of different time sequences in the same area, and obtain the anchor point change weight of each pane in each image according to the time sequence change of the pixel values ​​of different color channels at the center point of the pane; calculate the fluctuation of the grayscale value of the pixel points in the same sequence panes of adjacent time sequence images in the same area, construct a relationship function between turbidity and grayscale value fluctuation, and obtain the turbidity of each pane based on the fluctuation of the grayscale value of the pixel points in different panes; obtain the turbidity of each image based on the turbidity of each pane of each image in each area and the anchor point change weight of each pane of each image in each area, and calculate the average turbidity of all images in the same area as the turbidity of the area; assign the detection probability weight of the suspected garbage of the U-Net algorithm and the YOLO algorithm according to the turbidity of each area, and obtain the probability of the detection target being suspected garbage;

[0043] The pollutant identification module is used to set a probability threshold K, compare the probability threshold K with the probability that the detection target in each area is garbage, determine whether the detection target is garbage, and remind maintenance personnel to take corresponding measures.

[0044] A third aspect of the present invention is a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the above-mentioned method for intelligent identification of underwater pollutants based on image processing.

[0045] In a fourth aspect of the present invention, a computer device comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above-mentioned method for intelligent identification of underwater pollutants based on image processing when executing the computer program.

[0046] The beneficial effects of the technical solution of the present invention are:

[0047] Through time-series image acquisition, dual-algorithm probabilistic fusion, and turbidity adaptive weighting, the limitations of a single algorithm in turbid / clear waters are overcome, significantly improving the robustness and accuracy of pollutant identification.

[0048] Underwater images are collected from different areas at different times and are classified and archived according to the order of the image areas and time series. Interval-based shooting and a dual sorting mechanism ensure that the images cover multiple areas and are time-aligned, providing a structured data foundation for subsequent turbidity calculations and reducing motion blur interference.

[0049] Based on the pollutant recognition logic of the U-Net and YOLO algorithms, the detection locations of suspected garbage in the same image, as well as the U-Net and YOLO detection probabilities, are obtained from the two algorithms. The U-Net probability output is aligned with the YOLO detection frame as a reference, unifying the target location and probability of detection by the two algorithms to avoid missed detections and provide consistent data input for the fusion strategy.

[0050] Set up and traverse image panes of different time sequences in the same area. Based on the temporal changes in the pixel values ​​of different color channels at the center of the pane, obtain the anchor point change weight of each pane in each image. Calculate the anchor point change parameters by temporal difference of the RGB channels at the center of the pane. Combined with the regional mean normalization weight, this effectively suppresses the impact of instantaneous interference such as animal and plant occlusion on turbidity assessment.

[0051] The grayscale value fluctuations of pixels in the same sequence of adjacent time-series images in the same region are calculated, and a function is constructed to determine the relationship between turbidity and grayscale value fluctuations. The turbidity of each window is then determined by combining the grayscale value fluctuations of pixels in different window frames. A nonlinear mapping between grayscale fluctuations and turbidity is then established based on an exponential function, which better reflects the actual optical scattering characteristics and improves the physical rationality of the turbidity parameter c.

[0052] Based on the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region, the turbidity of each image is obtained. The average turbidity of all images in the same region is calculated as the weighted sum of the turbidity of the region, the pane turbidity, and the anchor point weight. This highlights the contribution of low-interference areas and improves the reliability of the regional turbidity W calculation.

[0053] The detection probability weights of the U-Net and YOLO algorithms for suspected garbage are assigned based on the turbidity level of each area. This yields the probability of detecting a target suspected of being garbage. The probabilities of the two algorithms are dynamically fused using the turbidity W as the weight, strengthening U-Net in turbid areas and prioritizing YOLO in clear areas, maximizing the complementary advantages of the algorithms.

[0054] A probability threshold K is set and compared with the probability of the detected target being garbage in each area to determine whether the detected target is garbage and remind maintenance personnel to take corresponding measures. The flexible setting of the probability threshold K meets the accuracy requirements of different water areas, reduces the false alarm rate, guides maintenance personnel to clean accurately, and improves operation and maintenance efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0056] Figure 1 This is a flowchart of the steps of an underwater pollutant intelligent identification method based on image processing according to the present invention;

[0057] Figure 2 This is a structural block diagram of an underwater pollutant intelligent identification system based on image processing according to the present invention. DETAILED DESCRIPTION

[0058] To further illustrate the technical means and effectiveness of the present invention in achieving its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, describes in detail the specific implementation, structure, features, and effectiveness of the image processing-based intelligent underwater contaminant identification method and system proposed by the present invention. In the following description, references to different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0059] Unless defined otherwise, 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 invention belongs.

[0060] The specific scheme of the underwater pollutant intelligent identification method and system based on image processing provided by the present invention is described in detail below with reference to the accompanying drawings.

[0061] See also Figure 1 , which shows the first object of the present invention, a flowchart of a method for intelligent identification of underwater pollutants based on image processing, the method comprising the following steps:

[0062] Step S001: Collect underwater images from different areas and at different time periods, and classify and archive the images according to the order of the image collection areas and the time sequence.

[0063] Since the existing single visual algorithm cannot adapt to turbid areas and clear areas at the same time, this method aims to improve the accuracy of underwater pollutant identification by calculating regional turbidity and combining two complementary visual algorithms. The turbidity in different areas may be different, so this step requires collecting images of different areas; and the collected images may be obstructed by plants and animals, resulting in inaccurate turbidity calculation, so it is necessary to continuously collect images of the same area to eliminate the influence of plant occlusion.

[0064] Specifically, underwater images of different areas and time periods are collected, and the images are classified and archived according to the order of the image areas and time series. The specific method is as follows:

[0065] An underwater camera is fixed one meter above the top of the underwater robot. The robot operates in an intermittent motion: after advancing U meters, it stops and captures an image every k seconds while stopped, for a total of T images. After capturing T images, the robot advances U meters again and repeats the process of stopping and capturing until it has captured images of S different areas, each with a size of L x L pixels. Ultimately, all captured images are doubly sorted and traversed, based on the sequence of areas corresponding to the robot's route and the chronological order of image acquisition within the same area.

[0066] It should be noted that in this embodiment, the forward distance U = 10 meters, the robot's travel route is a straight line, the time interval for taking images is v = 2 seconds, the number of images taken in each area is T = 20, and the number of areas corresponding to the collected images is S = 50. The present invention does not limit the robot's forward distance U, time interval v, number of shots T, and number of areas S each time. The specific selected values ​​depend on the specific implementation situation.

[0067] Step S002: Based on the pollutant recognition logic of the U-Net algorithm and the YOLO algorithm, the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability are respectively obtained by the two algorithms; image panes of different time sequences in the same area are set and traversed, and the anchor point change weight of each pane in each image is obtained according to the time sequence change of the pixel values ​​of different color channels at the center point of the pane; the fluctuation of the grayscale value of pixels in the same sequence panes of adjacent time sequence images in the same area is calculated, and a relationship function between turbidity and grayscale value fluctuation is constructed. The turbidity of each pane is obtained based on the fluctuation of the grayscale value of pixels in different panes; the turbidity of each image is obtained based on the turbidity of each pane in each image in each area and the anchor point change weight of each pane in each image in each area, and the average turbidity of all images in the same area is calculated as the turbidity of the area; the detection probability weights of the suspected garbage of the U-Net algorithm and the YOLO algorithm are assigned according to the turbidity of each area, and the probability of detecting the target as suspected garbage is obtained.

[0068] It should be noted that the commonly used algorithms for underwater robot recognition are the U-Net algorithm and the YOLO algorithm. U-Net, with its pixel-level segmentation capability, can accurately delineate polluted areas and provide detailed spatial information for pollution level assessment, while YOLO, with its high-speed target detection advantage, can quickly locate and frame discrete polluted targets, enabling real-time discovery and tracking of pollution sources. However, optical scattering caused by suspended particles in turbid areas significantly reduces image contrast and clarity, causing the performance of the YOLO algorithm, which relies on regular shapes and clear edge features, to decline. U-Net, based on pixel-level analysis, can effectively capture blurred targets through probabilistic output and an attention mechanism. Conversely, in clear areas, YOLO's efficient anchor box mechanism can fully utilize its advantages of fast detection speed and sensitivity to small targets, while U-Net may be inefficient in such environments due to computational redundancy. Therefore, the proposed fusion strategy based on regional turbidity level can combine the advantages of the two algorithms, improve robustness in complex areas, and enhance the recognition accuracy of underwater debris. This dynamic resource allocation can significantly improve the overall system's operating efficiency. Therefore, this step first requires the simultaneous extraction of the probabilistic outputs of the two algorithms.

[0069] It's also worth noting that both the U-Net and YOLO algorithms can independently identify underwater debris in different water quality environments (clear or turbid), but their performance is complementary: U-Net is better at accurately outlining debris through pixel-level segmentation in turbid areas, while YOLO excels at quickly locating discrete debris targets in clear areas. To combine the advantages of both algorithms, this study proposes a dynamic fusion mechanism. First, the probability of a detected target being trash, as output by U-Net, is obtained, along with the probability of the same target being trash detected by YOLO. Then, through an adaptive weighted fusion strategy, a more robust trash identification result is generated. To build a robust fusion model, the probabilistic outputs of both algorithms must be extracted simultaneously.

[0070] Specifically, based on the identification logic of the U-Net algorithm and the YOLO algorithm for pollutants, the detection locations of suspected garbage in the same image, as well as the U-Net detection probability and the YOLO detection probability, are obtained by the two algorithms respectively. The specific method is as follows:

[0071] For the same image, two algorithms are used to detect the image separately. When the YOLO algorithm successfully detects the target, the position of each YOLO recognition frame is output as the detection position of the suspected garbage. The probability that the target in the recognition frame is garbage is recorded as the detection probability of the YOLO algorithm in the recognition frame. ; Use the U-Net algorithm to calculate the mean probability of all pixels in each YOLO recognition frame being garbage, and use it as the U-Net algorithm detection probability that the target in the corresponding recognition frame of the image is garbage If the YOLO algorithm fails to detect the target, that is, there is no recognition box, the maximum value of the probability that the pixel point in the U-Net full image output is garbage is selected as the U-Net algorithm detection probability that there is garbage in the corresponding area of ​​the image. , and the location of the pixel is used as the detection location of suspected garbage.

[0072] It's important to note that both the U-Net and YOLO algorithms can independently identify underwater debris in different water quality environments (clear or turbid), but their performance is complementary: U-Net is better at accurately outlining debris through pixel-level segmentation in turbid areas, while YOLO excels at quickly locating discrete debris targets in clear areas. To combine the advantages of both algorithms, this study proposes a dynamic fusion mechanism—first, the garbage pixel probability map output by U-Net and the confidence level of the garbage targets detected by YOLO are separately obtained. Then, through an adaptive weighted fusion strategy, a more robust garbage identification result is generated.

[0073] It should be further explained that the above steps obtain the detection location of suspected garbage and the probability of the two algorithms judging the suspected garbage target as garbage. This step will judge the turbidity level based on the collected images, and then combine the turbidity level and the detection probability of different algorithms to obtain more accurate garbage identification results. However, the images obtained above are taken in time sequence, and aquatic plants and animals may enter the shooting screen, resulting in color mutations in the image and affecting the judgment of the turbidity level. Therefore, this step needs to reduce the impact of such situations on turbidity judgment.

[0074] Specifically, set and traverse the image panes of different time sequences in the same area, and obtain the anchor point change weight of each pane in each image based on the time sequence change of the pixel values ​​of different color channels at the center of the pane. The specific method is as follows:

[0075] Divide each L*L image into N*N panes. Starting from the upper left corner of the image, traverse each pane of the image in a snake-like manner to the right. For the nth pane of the tth image in the Sth region, calculate the pixel value change of the center pixel of the pane and the pixel value change of the center pixel of the same sequence pane in the previous image, and obtain the anchor point change parameter of the nth pane of the tth image in the Sth region. , the specific method is as follows:

[0076]

[0077] Where, The R channel pixel value of the center point of the nth pane of the tth image in the sth region, is the size of the G channel of the center point of the nth pane of the tth image in the sth region, is the size of the B channel of the center point of the nth pane of the tth image in the sth region, is the R channel pixel value of the center point of the nth pane of the t-1th image in the sth region, is the size of the G channel of the center point of the nth pane of the t-1th image in the sth region, is the size of the B channel of the center point of the nth pane of the t-1th image in the sth region, represents the anchor point change parameter of the nth pane in the tth image of the sth region;

[0078] It should be noted that, in this embodiment, the image side length is L = 1200 pixels, and the window pane side length is N = 120 pixels. The present invention does not limit the image side length L and the window pane side length N, but it must be ensured that L can be divided by N. In other embodiments, the image side length L and the window pane side length N depend on the specific implementation.

[0079] Calculate the mean of the anchor point change parameters for all panes of the same sequence in different images of the same region, and record the mean as the anchor point change weight coefficient for all panes of the same sequence in different images of the same region. The specific method is as follows:

[0080]

[0081] represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, represents the anchor point change parameter of the nth pane in the ith image of the sth region, represents the anchor point change weight coefficient of the nth window in the t+1th image of the sth region, where T represents the number of images collected for each region;

[0082] The above steps obtain the anchor point change weight coefficient of each pane. The anchor point change weight coefficient is calculated based on the mean of the anchor point change parameters of the panes of the same sequence in different images in the same area. Therefore, this step assigns a weight to each pane in each image in each area based on the size of the anchor point change weight coefficient. The pane with a larger anchor point change weight coefficient is assigned a smaller weight. Because when the anchor point change weight coefficient is large, it means that there is a higher probability that animals or suspended garbage exist in the pane corresponding to the anchor point change weight coefficient. The fluctuation of pixel values ​​caused by the turbid area is much smaller than the fluctuation of pixel values ​​caused by the movement of animals or suspended garbage. This may cause the calculation result of the method of calculating the turbidity degree based on pixel value fluctuation to be inaccurate. Therefore, the pane should be assigned a smaller weight, and the pane with a smaller anchor point change weight coefficient should be assigned a larger weight.

[0083] Specifically, according to the anchor point change weight coefficient of each pane of each image in each region, the anchor point change weight of each pane of each image in each region is obtained. The specific method is as follows:

[0084]

[0085] Where, represents the anchor point change weight of the nth pane of the tth image in the sth region, L represents the side length of the image, N represents the side length of the pane, represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, Represents the anchor point change weight coefficient of the j-th pane in the t-th image of the s-th region.

[0086] It should be noted that is based on The simplified result is to convert the relative size of the weight into weight, and perform normalization operation, and satisfy the following conditions: the larger the anchor point change weight coefficient, the smaller the pane weight, and the smaller the anchor point change weight coefficient, the larger the pane weight.

[0087] It should be further explained that the above steps have calculated the weight for each pane. The subsequent steps will calculate the turbidity of the corresponding area of ​​the entire image by combining the turbidity of the area corresponding to each pane with the pane anchor point change weight. The turbidity of each area is further obtained by the calculated value of the turbidity of the corresponding area of ​​all images in each area, and the change in the grayscale value of the pixel point in the same pane of adjacent images in the same area is judged. If the degree of change is large, it is considered that the turbidity may be high, otherwise it is considered that the turbidity may be low.

[0088] Specifically, the fluctuation of the grayscale values ​​of pixels in the same sequence panes of adjacent time-series images in the same area is calculated, and a function is constructed to determine the relationship between turbidity and grayscale fluctuation. The turbidity of each pane is obtained by combining the fluctuation of the grayscale values ​​of pixels in different panes. The specific method is as follows:

[0089] The average of the absolute values ​​of the differences between the grayscale values ​​of each pixel in the nth window pane of the tth image and the grayscale values ​​of each pixel in the nth window pane of the t-1th image in the same region is recorded as the turbidity parameter of the nth window pane of the tth image in the region. The turbidity parameter of each window pane in the first image is replaced by the average of the turbidity parameters of the window panes in the same sequence of each image in the same region.

[0090] The relationship function between the gray value change and the turbidity degree is established. The specific formula is as follows:

[0091]

[0092] Where g represents the degree of turbidity, represents the hyperparameter, c represents the turbidity parameter, substitute the turbidity parameter of the nth window pane of the tth image in the sth region into the relationship function to obtain the turbidity of the nth window pane of the tth image in the sth region. , represents the base of the natural logarithm, and further obtains the turbidity level of each pane of each image in each area;

[0093] It should be noted that the change in grayscale value and the degree of turbidity are not a simple linear relationship. The more obvious the change in grayscale value, the higher the degree of turbidity. However, when the turbidity exceeds the critical threshold, although the continued increase in turbidity will still expand the grayscale difference, the rate of increase will decay sharply (diminishing marginal effect). At the same time, the sensitivity of the visual algorithm accuracy to turbidity changes decreases at the same rate - because the suspended particles in highly turbid water bodies approach saturation concentration, the light scattering enhancement is limited, and the underwater visibility exceeds the algorithm's resolvable lower limit, causing the texture features to tend to be homogenized. At this time, the further increase in turbidity has little effect on the recognition accuracy. Generally speaking, the change between the previous and next frames in clear water will not exceed 20, while the difference between the previous and next frames of turbid liquids is above 60. Here, if the change in grayscale value between the previous and next frames exceeds 80, the probability of defining the liquid as turbid is very high. In this embodiment, the hyperparameter The value is 0.03. When the hyperparameter is 0.03, the correspondence between the grayscale change and the turbidity degree is more consistent with the theoretical model. The present invention does not limit the value of the hyperparameter, and the specific value of the hyperparameter depends on the implementation situation.

[0094] It should be further explained that the above process obtains the turbidity level of each pane of each image in each area, as well as the anchor point change weight of each pane of each image in each area. The turbidity level of each pane and the anchor point change weight of each pane are now combined to obtain the turbidity level of each image, and the average turbidity level of all images in the same area is further calculated as the turbidity level of the area.

[0095] Specifically, based on the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region, the turbidity of each image is obtained, and the average turbidity of all images in the same region is calculated as the turbidity of the region. The specific method is as follows:

[0096]

[0097] Where L represents the image side length, N represents the grid side length, represents the turbidity of the t-th image in the s-th region, represents the turbidity of the I-th window pane of the t-th image in the s-th region, represents the anchor point change weight of the I-th pane of the t-th image in the s-th region;

[0098] Calculate the average turbidity of all images in each region as the turbidity of each region. The specific method is as follows:

[0099]

[0100] Where T represents the number of all images in a region, represents the turbidity level of the sth region, Indicates the turbidity level of the J-th image in the s-th region.

[0101] It should be noted that the above method determines the turbidity of each image based on the turbidity and weight of each area in the image, and then calculates the turbidity of each area based on the turbidity of different images in each area. Subsequently, the U-Net algorithm detection results and the YOLO algorithm detection results are assigned weights according to the size of the turbidity of the area to obtain the probability that the detection target is garbage.

[0102] Specifically, the detection probability weights of the U-Net algorithm and the YOLO algorithm are assigned according to the turbidity level of each area to obtain the probability of the detected target being suspected of being garbage. The specific method is as follows:

[0103]

[0104] Where, represents the probability of detecting a suspected garbage target in the sth region, represents the turbidity level of the sth region, It represents the probability of the U-Net algorithm detecting the suspected garbage in the sth region. It represents the probability of the YOLO algorithm detecting the suspected garbage when detecting the sth region.

[0105] It should be noted that by analyzing the change of the gray value of each grid in time series, the turbidity of each pane can be obtained. Then, according to the anchor point change weight of each pane and the turbidity calculated for each pane, the turbidity of each image can be obtained. Furthermore, the actual turbidity of each area can be obtained. Finally, the regional turbidity is used as the weight to adjust and The output proportion of is finally obtained to obtain the probability of judging the target as garbage. The actual turbidity level of the area can be used as a weight because its calculation method is based on the mean of the relationship function g, and the value range of the relationship function g is 0 to infinitely close to 1.

[0106] Step S003: Setting a probability threshold K, comparing the probability threshold K with the probability of the detected target being garbage in each area, determining whether the detected target is garbage and notifying maintenance personnel to perform corresponding processing.

[0107] Specifically, a probability threshold K is set, and the probability of the detected target being garbage in each area is compared with the probability threshold K to determine whether the detected target is garbage and to remind maintenance personnel to take corresponding measures. The specific method is as follows:

[0108] Based on the requirements for detection result accuracy in different waters, a probability threshold K is set and compared with the probability P of the detection target in each area being garbage. If P is greater than the probability threshold K, the detection target is determined to be garbage, and maintenance personnel are reminded to salvage and clean it up.

[0109] It should be noted that, through the above method, we judge the turbidity of the water area and provide dynamic weights for the probability-enhanced U-Net algorithm and the YOLO algorithm, thereby realizing an intelligent recognition method that can adapt to various water environments and accurately identify underwater garbage. In this embodiment, the probability threshold K=0.7. The present invention does not limit the value of the probability threshold K, and the specific value depends on the implementation situation.

[0110] See also Figure 2 , which shows the second object of the present invention, a structural block diagram of an underwater pollutant intelligent identification system based on image processing, the system includes the following modules:

[0111] Underwater image acquisition module, used to collect underwater images of different areas and time periods, and classify and archive the images according to the order of the image areas and time series;

[0112] The turbidity calculation module is used to obtain the detection position of suspected garbage in the same image and the U-Net detection probability and YOLO detection probability of the two algorithms based on the recognition logic of pollutants of the U-Net algorithm and the YOLO algorithm respectively; set and traverse the image panes of different time sequences in the same area, and obtain the anchor point change weight of each pane in each image according to the time sequence change of the pixel values ​​of different color channels at the center point of the pane; calculate the fluctuation of the grayscale value of the pixel points in the same sequence panes of adjacent time sequence images in the same area, construct a relationship function between turbidity and grayscale value fluctuation, and obtain the turbidity of each pane based on the fluctuation of the grayscale value of the pixel points in different panes; obtain the turbidity of each image based on the turbidity of each pane of each image in each area and the anchor point change weight of each pane of each image in each area, and calculate the average turbidity of all images in the same area as the turbidity of the area; assign the detection probability weight of the suspected garbage of the U-Net algorithm and the YOLO algorithm according to the turbidity of each area, and obtain the probability of the detection target being suspected garbage;

[0113] The pollutant identification module is used to set a probability threshold K, compare the probability threshold K with the probability that the detection target in each area is garbage, determine whether the detection target is garbage, and remind maintenance personnel to take corresponding measures.

[0114] The third object of an embodiment of the present invention is to provide a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the steps of the above-mentioned method for intelligent identification of underwater pollutants based on image processing are implemented.

[0115] The fourth object of an embodiment of the present invention is to provide a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned method for intelligent identification of underwater pollutants based on image processing are implemented.

[0116] This embodiment uses time-series image acquisition, dual-algorithm probabilistic fusion, and turbidity adaptive weighting to overcome the limitations of a single algorithm in turbid / clear waters, significantly improving the robustness and accuracy of pollutant identification.

[0117] Underwater images are collected from different areas at different times and are classified and archived according to the order of the image areas and time series. Interval-based shooting and a dual sorting mechanism ensure that the images cover multiple areas and are time-aligned, providing a structured data foundation for subsequent turbidity calculations and reducing motion blur interference.

[0118] Based on the pollutant recognition logic of the U-Net and YOLO algorithms, the detection locations of suspected garbage in the same image, as well as the U-Net and YOLO detection probabilities, are obtained from the two algorithms. The U-Net probability output is aligned with the YOLO detection frame as a reference, unifying the target location and probability of detection by the two algorithms to avoid missed detections and provide consistent data input for the fusion strategy.

[0119] Set up and traverse image panes of different time sequences in the same area. Based on the temporal changes in the pixel values ​​of different color channels at the center of the pane, obtain the anchor point change weight of each pane in each image. Calculate the anchor point change parameters by temporal difference of the RGB channels at the center of the pane. Combined with the regional mean normalization weight, this effectively suppresses the impact of instantaneous interference such as animal and plant occlusion on turbidity assessment.

[0120] The grayscale value fluctuations of pixels in the same sequence of adjacent time-series images in the same region are calculated, and a function is constructed to determine the relationship between turbidity and grayscale value fluctuations. The turbidity of each window is then determined by combining the grayscale value fluctuations of pixels in different window frames. A nonlinear mapping between grayscale fluctuations and turbidity is then established based on an exponential function, which better reflects the actual optical scattering characteristics and improves the physical rationality of the turbidity parameter c.

[0121] Based on the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region, the turbidity of each image is obtained. The average turbidity of all images in the same region is calculated as the weighted sum of the turbidity of the region, the pane turbidity, and the anchor point weight. This highlights the contribution of low-interference areas and improves the reliability of the regional turbidity W calculation.

[0122] The detection probability weights of the U-Net and YOLO algorithms for suspected garbage are assigned based on the turbidity level of each area. This yields the probability of detecting a target suspected of being garbage. The probabilities of the two algorithms are dynamically fused using the turbidity W as the weight, strengthening U-Net in turbid areas and prioritizing YOLO in clear areas, maximizing the complementary advantages of the algorithms.

[0123] A probability threshold K is set and compared with the probability of the detected target being garbage in each area to determine whether the detected target is garbage and remind maintenance personnel to take corresponding measures. The flexible setting of the probability threshold K meets the accuracy requirements of different water areas, reduces the false alarm rate, guides maintenance personnel to clean accurately, and improves operation and maintenance efficiency.

[0124] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0125] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1A device that provides the functions specified in a block or multiple blocks.

[0126] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0127] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0128] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered by the scope of protection of the claims of the present invention.

Claims

1. An intelligent identification method for underwater pollutants based on image processing, characterized in that: The method comprises the following steps: Collect underwater images from different areas and at different times, and classify and archive the images according to the order of the image collection areas and time series; According to the recognition logic of the U-Net algorithm and the YOLO algorithm for pollutants, the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability are respectively obtained by the two algorithms; the recognition logic of the U-Net algorithm and the YOLO algorithm for pollutants, the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability are respectively obtained by the two algorithms, including the specific method of: for the same image, using two algorithms to detect the image respectively, when the YOLO algorithm successfully detects the target, outputting the position of each YOLO recognition frame as the detection position of the suspected garbage, and the probability that the target in the recognition frame is garbage is recorded as the detection probability of the YOLO algorithm in the recognition frame. ; Use the U-Net algorithm to calculate the mean probability of all pixels in each YOLO recognition frame being garbage, and use it as the U-Net algorithm detection probability that the target in the corresponding recognition frame of the image is garbage If the YOLO algorithm fails to detect the target, that is, there is no recognition box, the maximum value of the probability that the pixel point in the U-Net full image output is garbage is selected as the U-Net algorithm detection probability that there is garbage in the corresponding area of ​​the image. , and use the location of the pixel as the detection location of suspected garbage; set and traverse the image panes of different time series in the same area, and obtain the anchor point change weight of each pane in each image according to the time series change of the pixel values ​​of different color channels at the center point of the pane; calculate the fluctuation of the grayscale value of the pixel points in the same sequence panes of adjacent time series images in the same area, build a relationship function between turbidity and grayscale value fluctuation, and obtain the turbidity of each pane based on the fluctuation of the grayscale value of the pixel points in different panes; obtain the turbidity of each image according to the turbidity of each pane of each image in each area and the anchor point change weight of each pane of each image in each area, and calculate the average turbidity of all images in the same area as the turbidity of the area; assign the detection probability weight of the suspected garbage of the U-Net algorithm and the YOLO algorithm according to the turbidity of each area, and obtain the probability of detecting the target as suspected garbage; A probability threshold K is set, and the probability of the detected target being garbage in each area is compared with the probability threshold K to determine whether the detected target is garbage and remind maintenance personnel to take corresponding measures.

2. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The method of collecting underwater images of different areas and different time periods and classifying and archiving the images according to the order of the collected image areas and the time sequence includes the following specific methods: An underwater camera is fixed one meter above the top of the underwater robot. The robot works in an intermittent movement mode: it stops after advancing U meters each time and takes an image of the current area every k seconds while stopped, for a total of T images of the current area. After completing T images, the robot advances U meters again and repeats the stopping process until it completes the image acquisition of S different areas, with an image size of L*L pixels. Finally, all the acquired images will be doubly sorted and traversed according to the area sequence corresponding to the robot's route and the time order of image acquisition within the same area.

3. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The method of setting and traversing image panes of different time sequences in the same area and obtaining the anchor point change weight of each pane in each image according to the time sequence change of pixel values ​​of different color channels at the center of the pane includes the following specific methods: Divide each L*L image into N*N panes. Starting from the upper left corner of the image, traverse each pane of the image in a snake-like manner to the right. For the nth pane of the tth image in the Sth region, calculate the pixel value change of the center pixel of the pane and the pixel value change of the center pixel of the same sequence pane in the previous image, and obtain the anchor point change parameter of the nth pane of the tth image in the Sth region. , the specific method is as follows: Where, The R channel pixel value of the center point of the nth pane of the tth image in the sth region, is the size of the G channel of the center point of the nth pane of the tth image in the sth region, is the size of the B channel of the center point of the nth pane of the tth image in the sth region, is the R channel pixel value of the center point of the nth pane of the t-1th image in the sth region, is the size of the G channel of the center point of the nth pane of the t-1th image in the sth region, is the size of the B channel of the center point of the nth pane of the t-1th image in the sth region, represents the anchor point change parameter of the nth pane in the tth image of the sth region; Calculate the mean of the anchor point change parameters for all panes of the same sequence in different images of the same region, and record the mean as the anchor point change weight coefficient for all panes of the same sequence in different images of the same region. The specific method is as follows: represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, represents the anchor point change parameter of the nth pane in the ith image of the sth region, represents the anchor point change weight coefficient of the nth window in the t+1th image of the sth region, where T represents the number of images collected for each region; According to the anchor point change weight coefficient of each pane of each image in each region, the anchor point change weight of each pane of each image in each region is obtained as follows: Where, represents the anchor point change weight of the nth pane of the tth image in the sth region, L represents the side length of the image, N represents the side length of the pane, represents the anchor point change weight coefficient of the nth pane in the tth image of the sth region, Represents the anchor point change weight coefficient of the j-th pane in the t-th image of the s-th region.

4. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The method of calculating the fluctuation of the grayscale values ​​of pixels in the same sequence panes of adjacent time-series images in the same area, constructing a function relating the degree of turbidity and the fluctuation of the grayscale values, and obtaining the degree of turbidity of each pane in combination with the fluctuation of the grayscale values ​​of pixels in different panes includes the following specific methods: The average of the absolute values ​​of the differences between the grayscale values ​​of each pixel in the nth window pane of the tth image and the grayscale values ​​of each pixel in the nth window pane of the t-1th image in the same region is recorded as the turbidity parameter of the nth window pane of the tth image in the region. The turbidity parameter of each window pane in the first image is replaced by the average of the turbidity parameters of the window panes in the same sequence of each image in the same region. The relationship function between the gray value change and the turbidity degree is established. The specific formula is as follows: Where g represents the degree of turbidity, represents the hyperparameter, c represents the turbidity parameter, substitute the turbidity parameter of the nth window pane of the tth image in the sth region into the relationship function to obtain the turbidity of the nth window pane of the tth image in the sth region. , Representing the base of the natural logarithm, we further obtain the turbidity level of each pane of each image in each region.

5. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The method of obtaining the turbidity of each image based on the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region and calculating the average turbidity of all images in the same region as the turbidity of the region includes the following specific methods: Where L is the side length of the image, N is the side length of the pane, represents the turbidity of the t-th image in the s-th region, represents the turbidity of the I-th window pane of the t-th image in the s-th region, represents the anchor point change weight of the I-th pane of the t-th image in the s-th region; Calculate the average turbidity of all images in each region as the turbidity of each region. The specific method is as follows: Where T represents the number of all images in a region, represents the turbidity level of the sth region, Indicates the turbidity level of the J-th image in the s-th region.

6. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The specific method for allocating the detection probability weights of the suspected garbage by the U-Net algorithm and the YOLO algorithm according to the turbidity level of each area and obtaining the probability of detecting the target suspected garbage is as follows: Where, represents the probability of detecting a suspected garbage target in the sth region, represents the turbidity level of the sth region, It represents the probability of the U-Net algorithm detecting the suspected garbage in the sth region. It represents the probability of the YOLO algorithm detecting the suspected garbage when detecting the sth region.

7. The method for intelligent identification of underwater pollutants based on image processing according to claim 1, characterized in that: The method of setting a probability threshold K and comparing the probability threshold K with the probability of the detected target in each area being garbage determines whether the detected target is garbage and reminds maintenance personnel to take corresponding measures. The specific method includes: Based on the requirements for detection result accuracy in different waters, a probability threshold K is set and compared with the probability P of the detection target in each area being garbage. If P is greater than the probability threshold K, the detection target is determined to be garbage, and maintenance personnel are reminded to salvage and clean it up.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method for intelligent identification of underwater pollutants based on image processing as described in any one of claims 1 to 7 are implemented.

9. An intelligent underwater pollutant identification system based on image processing, characterized in that: The system includes the following modules: Underwater image acquisition module, used to collect underwater images of different areas and time periods, and classify and archive the images according to the order of the image areas and time series; The turbidity calculation module is used to obtain the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability respectively according to the U-Net algorithm and the YOLO algorithm's recognition logic for pollutants; the specific method of obtaining the detection position of suspected garbage in the same image and the U-Net detection probability and the YOLO detection probability respectively according to the U-Net algorithm and the YOLO algorithm's recognition logic for pollutants is as follows: for the same image, use two algorithms to detect the image respectively, when the YOLO algorithm successfully detects the target, output the position of each YOLO recognition frame as the detection position of the suspected garbage, and the probability that the target in the recognition frame is garbage is recorded as the detection probability of the YOLO algorithm in the recognition frame. ; Use the U-Net algorithm to calculate the mean probability of all pixels in each YOLO recognition frame being garbage, and use it as the U-Net algorithm detection probability that the target in the corresponding recognition frame of the image is garbage If the YOLO algorithm fails to detect the target, that is, there is no recognition box, the maximum value of the probability that the pixel point in the U-Net full image output is garbage is selected as the U-Net algorithm detection probability that there is garbage in the corresponding area of ​​the image. , and use the location of the pixel as the suspected garbage detection location; set and traverse the image panes of different time sequences in the same area, and obtain the anchor point change weight of each pane in each image based on the time sequence change of the pixel values ​​of different color channels at the center of the pane; Calculate the fluctuation of the grayscale values ​​of pixels in the same sequence panes of adjacent time-series images in the same area, construct a function to relate the turbidity level to the grayscale value fluctuation, and obtain the turbidity level of each pane by combining the fluctuation of the grayscale values ​​of pixels in different panes; According to the turbidity of each pane of each image in each region and the anchor point change weight of each pane of each image in each region, the turbidity of each image is obtained, and the average turbidity of all images in the same region is calculated as the turbidity of the region; Assign the detection probability weights of suspected garbage by the U-Net algorithm and the YOLO algorithm based on the turbidity level of each area to obtain the probability of the detected target being suspected garbage; The pollutant identification module is used to set a probability threshold K, compare the probability threshold K with the probability that the detection target in each area is garbage, determine whether the detection target is garbage, and remind maintenance personnel to take corresponding measures.

Images