Image processing and recognition method for assisting underwater robot operation

Abstract

The application discloses an image processing and recognition method for assisting underwater robot operation and belongs to the technical field of underwater robot operation image processing. In order to solve the problems of image blurring and low work efficiency of the existing underwater robot image processing system, the application combines the characteristics that the light source of the image returned by the underwater robot is an artificial light source, divides the image into a light source area and a non-light source area, brings the foggy dark channel image after removing the light source into an atmospheric light component algorithm to obtain atmospheric light component A1, brings the original foggy dark channel image into the atmospheric light component algorithm to obtain atmospheric light component A0, then performs weighted fusion on A0 and A1, and further obtains a defogging image with better exposure. The application replaces downsampling of YOLOV5, and improves training set cleaning. The better defogging image is input into the improved YOLOV5 model for target recognition, so that the recognition accuracy can be effectively improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing of underwater robot operation, and relates to an image processing method and an identification method for underwater robot operation. BACKGROUND

[0002] An underwater robot is a robot capable of detecting, exploring and operating underwater. It has strong autonomous control ability and adaptability, and can work in various water environments such as oceans, rivers and lakes. However, the underwater environment is more complex than the terrestrial environment, and the complex environmental factors such as insufficient light, blurred vision and turbulent water flow cause great difficulties in underwater search. Many problems still need to be solved.

[0003] At present, many scholars have studied the solutions to the problem of underwater image blur. The solutions can be mainly divided into three types of algorithms: general image processing algorithms, algorithms based on imaging physical models and deep learning algorithms. Although the general image processing algorithm is very mature, its applicability is poor. The algorithm based on the imaging physical model performs well but is still in the development stage and is not mature. The deep learning algorithm can obtain an algorithm with low time consumption and good performance under the condition of a high-quality data set, but the number of high-quality data sets of underwater images is small. SUMMARY

[0004] The application is to solve the problems of image blur and low work efficiency of the existing underwater robot image processing system.

[0005] An image processing method for assisting underwater robot operation, comprising the following steps:

[0006] Step 1, collecting a video image by using an underwater robot camera, denoted as an original image;

[0007] Step 2, processing the video image by using an improved DCP algorithm, wherein the DCP algorithm is a dark channel prior dehazing algorithm;

[0008] The process of processing the video image by using the improved DCP algorithm comprises the following steps:

[0009] An original foggy dark channel image is obtained, denoted as an original foggy dark channel image;

[0010] For the original foggy dark channel image, first, a region segmentation method is used to distinguish a light source area and a non-light source area, then a binary image of the non-light source area is extracted and multiplied with the foggy dark channel image to obtain a foggy dark channel image after removing the light source; and the foggy dark channel image after removing the light source is brought into an atmospheric light component algorithm to obtain an atmospheric light component A1.

[0011] The original foggy dark channel image is brought into the atmospheric light component A0 obtained by the atmospheric light component algorithm, that is, the atmospheric light component intensity of the light source region;

[0012] The A0 and A1 are weighted and fused to obtain the final atmospheric light component A2, and the specific mathematical expression is as follows:

[0013]

[0014] Wherein, x represents the pixel position, y1 and y2 represent the pixel position of the light source region and the pixel position of the non-light source region; Alpha represents the fusion coefficient; Ar0(x) and Ar1(x) represent the light source region and the non-light source region; N0 and n1 represent the total number of pixel points of the light source region and the total number of pixel points of the non-light source region; The fusion coefficient alpha is as follows:

[0015]

[0016] A 0_1 = In(1+A0)

[0017] A 1_1 = In(1+A1)

[0018] Wherein, beta is a correction coefficient; A 0_1 is the light source region part of the atmospheric light component after processing the atmospheric light component A0, A 1_1 is the non-light source region part of the atmospheric light component after processing the atmospheric light component A1.

[0019] The transmittance is calculated by using the light source region atmospheric light component A0:

[0020]

[0021] Wherein, w is a constant parameter representing the reserved depth, I c (y) is a foggy image of a color channel; Omega(x) is a square region with x as the center.

[0022] The dehazing image is obtained by using A2:

[0023]

[0024] Wherein, t0 represents the lower limit of transmittance; I(x) represents the foggy image.

[0025] Preferably, the correction coefficient beta is 0.7.

[0026] Preferably, the constant parameter w representing the reserved depth is 0.95.

[0027] Preferably, the lower limit of transmittance t0 is 0.1.

[0028] Preferably, the process of calculating transmittance by using the light source region atmospheric light component A0 is replaced by the following process:

[0029]

[0030] Wherein, w is a constant parameter representing the depth of field, I dark (x) represents a foggy dark channel image.

[0031] Preferably, in the process of estimating the atmospheric light component, the image resolution is compressed by n times downsampling of the foggy dark channel image, reducing the data volume of the foggy dark channel image to 1 / n 2 of the original, and then sorting all pixel points of the foggy dark channel image, and estimating the atmospheric light component.

[0032] Preferably, in the process of processing the video image by using the improved DCP algorithm, the original image after grayscale is used as a guide image to guide the filtering of the estimated transmittance, and for the case that a large number of mean filters are required for image processing in the guided filtering algorithm, GPU parallel computing method is used to process the image by multiple mean filters at the same time.

[0033] Preferably, the image in the guided filtering adopts 64-bit floating point data, and the images in other processes all adopt 32-bit floating point data.

[0034] An image recognition method for assisting underwater robot operation, comprising the following steps:

[0035] First, the video image collected by the underwater robot camera is processed by using the image processing method for assisting underwater robot operation; then the target is recognized by using the improved YOLOV5 algorithm.

[0036] In the improved YOLOV5 algorithm, YOLOV5s is selected as the basic model, and the pyramid structure combining FPN and PAN is used in the neck of YOLOV5 for feature enhancement; three times of up-sampling are required for the image before inputting the pyramid structure, and deconvolution is used for up-sampling during the three times of up-sampling.

[0037] Preferably, the improved YOLOV5 algorithm needs to be cleaned before training, and the improved YOLOV5 algorithm is trained by using the cleaned data set, so as to obtain the trained improved YOLOV5 algorithm.

[0038] The process of data set cleaning comprises the following steps:

[0039] First, randomly extract pictures from the COCO dataset; Then, filter out only the selected target classification labels and pictures from the randomly extracted pictures to obtain the pre-cleaning dataset; The label of the pre-cleaning dataset is manually checked, and the label frame is modified and improved, that is, the COCO cleaning dataset for training the improved YOLOV5 algorithm is obtained.

[0040] Beneficial effects:

[0041] The DCP algorithm and the YOLOV5 recognition algorithm are improved, wherein the improved DCP algorithm is improved by improving the atmospheric light component algorithm to realize performance optimization, so that the contrast of the image returned by the underwater robot is improved, which is beneficial to the operator to observe the image; In addition, the output image of the DCP is input to the improved target recognition algorithm, the YOLOV5 target recognition algorithm improves the upsampling mode, and the data set cleaning process is also improved, which is beneficial to the neural network to extract image features, so as to alleviate the problem of lack of underwater image training set of deep learning algorithm, improve the target recognition accuracy, assist the operator to locate the searched object faster, and improve the operation efficiency of the underwater robot.

[0042] The GPU parallel acceleration, downsampling compression data, algorithm simplification and mixed precision mode can greatly reduce the time consumption, and can further improve the operation efficiency of the underwater robot. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 The figure is a schematic diagram of an image processing process for assisting underwater robot operation. DETAILED DESCRIPTION

[0044] To solve the image blur problem and improve the work efficiency, an improved dark channel defogging algorithm and an improved YOLOV5 target recognition algorithm are proposed. The dark channel defogging algorithm (DCP) based on the imaging physical model is improved in terms of image quality. In addition, to improve the efficiency of underwater operation of the underwater robot, GPU parallel acceleration, downsampling compression data, algorithm simplification and mixed precision mode are used to reduce the time consumption of the DCP algorithm. The target recognition algorithm is replaced by upsampling and data set cleaning to improve the recognition accuracy.

[0045] Specific implementation method one: combining Figure 1 It is explained that the embodiment,

[0046] The embodiment is an image processing method for assisting underwater robot operation, comprising the following steps:

[0047] Step 1, using the underwater robot camera to collect video images, denoted as original images;

[0048] Step 2, the video image is processed by using improved DCP algorithm:

[0049] To solve the problem of light scattering by particles, Narasimhan et al. proposed an atmospheric scattering physical model, which provides a theoretical basis for restoring the original image from the fog image on the mathematical level. The mathematical expression of the atmospheric scattering model is shown in equation (1).

[0050] I(x) = J(x)t(x) + A(1-t(x)) (1)

[0051] Where I(x) represents the fog image; J(x) represents the non-fog image; t(x) represents the transmittance; A represents the atmospheric light component.

[0052] De-fogging and de-dusting has always been a hot direction in the field of image processing, and the DCP algorithm has become a classic de-fogging algorithm with its simple method and good results. He et al. obtained the dark channel priori knowledge through a large number of landscape pictures, which made the solution of the atmospheric scattering model simple. The mathematical expression of the dark channel priori theory is shown in equation (2).

[0053]

[0054]

[0055] Where J dark (x) represents the non-fog dark channel image, I dark (x) represents the fog dark channel image; Ω(x) is a square region centered at x; c is the RGB three color channel; J c (y) is the non-fog image of the color channel, and I c (y) is the fog image of the color channel.

[0056] Generally, it can be assumed that the atmospheric light component A is known. By using the priori knowledge that one of the three channels tends to zero, equation (1) and equation (2) can be combined to obtain the estimated transmittance mathematical expression:

[0057]

[0058] Where w is a constant parameter with 0.95, indicating that the depth of field is retained; A c is the atmospheric light component of the color fog image (the parameter subscript is c The parameters without the subscript c are the same parameters as the parameters without the subscript c The main purpose is to emphasize that the calculation of the parameter with the subscript c requires R, G, and B channels to be calculated separately to obtain the result).

[0059] However, in the actual situation, the image environment is complex, and A is not a known quantity. In this case, the top 0.1% of the brightest pixels in the foggy dark channel image can be selected, and the average value can be obtained.

[0060] The foggy image of the color channel and the estimated atmospheric light component are brought into mathematical expression (4) to obtain the estimated transmittance. After a series of mathematical transformations of the atmospheric scattering model, the mathematical expression of the de-fogging image is obtained as shown in equation (5).

[0061]

[0062] where t0 is 0.1, representing the lower limit of transmittance, to avoid losing original image information when the transmittance is close to 0;

[0063] Although the DCP algorithm can better de-fog the image, the use of minimum filtering in the calculation of the foggy dark channel image leads to image edge blurring and white edges in the de-fogging image. Therefore, He further proposed a guided filtering algorithm that uses the grayscale original image as a guide to filter the estimated transmittance, making the estimated transmittance edge sharpened and the image more detailed, and the white edges in the de-fogging image disappear.

[0064] Directly applying the DCP algorithm to the underwater robot image processing can achieve the effect of increasing contrast and de-blurring, but the DCP algorithm will make the image overall dark, and the underwater image has a low exposure, so the brightness is very low after being processed by the DCP algorithm, and the image information is less, which is not conducive to the observation of the operator.

[0065] The atmospheric light component A in the sky is the light after the sun is scattered by the atmosphere, which is affected by the atmospheric concentration and the fog concentration. Since the light source of the image is the searchlight of the underwater robot, the brightness near the searchlight is high, so the top 0.1% of the brightest pixels in the foggy dark channel image is always at the light source. Therefore, the atmospheric light component A of the actual image of the underwater robot is a high constant, which leads to the inapplicability of the prior theory of the DCP algorithm, and the processed image effect is not ideal. Therefore, the algorithm for the atmospheric light component A is improved to adapt to the actual situation of the underwater robot.

[0066] Improved DCP algorithm: when improving the atmospheric light component A algorithm, firstly consider the influence of the robot searchlight, which needs to be distinguished from the scattering light source of the suspended particles to obtain the correct atmospheric light component, and the angle of the robot camera can be adjusted by the operator, which leads to the light source area is not a fixed area, so the method of region segmentation is adopted to distinguish the light source area and the non-light source area. Based on the obtained foggy dark channel image (denoted as the original foggy dark channel image), it can be observed that the pixel value of the light source is larger, and the pixel value of the remaining area is smaller, so after obtaining the global optimal threshold value by using the Otsu algorithm, the image is binarized according to the threshold value to realize the segmentation of the light source area and the non-light source area. The binarization can change the light source area into a region with pixel value of 0, and the non-light source area into a region with pixel value of 1. Then, the binarized image of the non-light source area is multiplied with the foggy dark channel image to obtain the foggy dark channel image after removing the light source.

[0067] The foggy dark channel image after removing the light source is brought into the atmospheric light component algorithm (the atmospheric light algorithm is an algorithm for obtaining the parameter A of the atmospheric scattering model) to obtain the atmospheric light component A1.

[0068] The original foggy dark channel image (the foggy dark channel image before light source segmentation) is brought into the atmospheric light component algorithm to obtain the atmospheric light component A0, which is the atmospheric light component intensity of the light source area.

[0069] Unlike the light scattering in the sky, the light scattering in water is more serious, and the light intensity decay is larger, so A1 cannot be directly used as the parameter of the scattering model, and a certain amount of A0 value at the light source needs to be preserved. A0 and A1 are weighted and fused to obtain the final atmospheric light component A2, and the specific mathematical expression is shown in formula (6).

[0070]

[0071] Wherein, x represents the pixel position, y1 and y2 represent the pixel position of the light source area and the pixel position of the non-light source area. Alpha represents the fusion coefficient, which preserves a certain amount of light source brightness; Ar0(x) and Ar1(x) represent the light source area and the non-light source area; n0 and n1 represent the total number of pixel points in the light source area and the total number of pixel points in the non-light source area.

[0072] In order to obtain a strong adaptive fusion coefficient alpha, the atmospheric light components A0 and A1 are first preprocessed, so that the processed parameters are not sensitive to the change of A0 corresponding to the light source area part and sensitive to the change of A1 corresponding to the non-light source area part, and the specific mathematical expressions are shown in formulas (7) and (8):

[0073] A 0_1 = In(1+A0) (7)

[0074] A1_1 = In(1 + A1) (8)

[0075] where A 0_1 is the processed light source region part of the atmospheric light component, A 1_1 is the processed non-light source region part of the atmospheric light component.

[0076] The difference between the processed atmospheric light components can obtain the brightness difference between the light source region and the non-light source region. The fusion coefficient a should be proportional to the size of the brightness difference, that is, the darker the non-light source region, the larger the fusion coefficient, and the larger the proportion of the light source region in the atmospheric intensity A2, so as to achieve the effect of better image exposure. The specific expression is shown in (9).

[0077]

[0078] where β is a correction coefficient, so that the difference is in the interval of the independent variable with a suitable change rate of exponential function. The experimental results show that the effect is best when β is 0.7.

[0079] By analyzing the original DCP algorithm, the atmospheric light component value will affect the estimation of transmittance, guided filtering, and the solution of the mathematical expression of the defogging image. When calculating the estimated transmittance value, the approximate estimation of the ratio of the sky region to the atmospheric light component close to 1 is needed, and if A2 is used for approximate estimation, it will not be true, and the estimated transmittance is difficult to solve. In the actual underwater image, the light source is regarded as the sky region, so the atmospheric light component A0 of the light source region is still used when calculating the transmittance value. In the mathematical expression of the defogging image (formula (5)), the physical meaning of the atmospheric light component is the atmospheric light component at infinity in the atmospheric scattered light, so the above modified A2 value is used for model solving. The specific mathematical expressions of the improved algorithm are shown in (10) and (11):

[0080]

[0081]

[0082] Fast implementation of DCP algorithm: Since underwater robots need real-time operation, image processing algorithms applied to underwater robots have low time consumption requirements. Although the DCP algorithm has small computational complexity, it cannot achieve real-time effect, so the parts of the algorithm need to be analyzed and optimized to shorten the algorithm time consumption. The specific improvements are as follows:

[0083] (a) Down-sampling to compress data volume: When estimating the atmospheric light component, all pixel points of the foggy dark channel image need to be sorted, which results in large data volume and causes large time consumption even if a fast sorting algorithm is used. Therefore, the image resolution is compressed by n times down-sampling, which reduces the data volume of the foggy dark channel image to 1 / n of the original. 2That is, the time-consuming of the sorting algorithm can be reduced by n 2 times.

[0084] (b) Algorithm simplification: When calculating the estimated transmittance to solve the mathematical expression of the defogging image, the original image needs to be divided by A0 and then obtain the dark channel image again (this dark channel image does not refer to a specific meaning, only represents the min(min(·)) algorithm). Due to the actual situation, the pixel value of the light source area is large, which causes A0 of the three channels to be constant. After the original image is divided by the constant, the pixel point corresponding to the position of the minimum value of the three channels after division by the constant does not change, and the pixel point position of the corresponding dark channel image does not change. Therefore, it is not necessary to obtain the dark channel image again, and only the foggy dark channel image obtained in the first step needs to be substituted into the mathematical expression of the estimated transmittance to express the specific expression as formula (12):

[0085]

[0086] (c) GPU parallel acceleration: In the deep learning model, due to the large amount of convolution operation, GPU parallel is often used to make multiple convolutions operate at the same time to realize the acceleration of the algorithm. Referring to the deep learning acceleration method, for the case that a large number of mean filters are needed for image processing in the guided filtering algorithm, the GPU parallel computing method is used to make multiple mean filters process the image at the same time, which can reduce the time-consuming of the guided filtering algorithm.

[0087] (d) Mixed precision operation: In the original algorithm, in order to ensure that the details of the image are good, all data types are processed by 64-bit floating point type, while mixed precision calculation is often used in deep learning to reduce the amount of calculation and reduce the training time. Since the purpose of guided filtering is to increase the details of the image, the image in the guided filtering uses 64-bit floating point type, while the rest of the images use 32-bit floating point type to ensure the quality of the image while reducing the amount of calculation. Specific implementation method two:

[0089] The image recognition method for assisting the operation of the underwater robot includes the following steps:

[0090] First, the video image collected by the underwater robot camera is processed by using the image processing method for assisting the operation of the underwater robot described in the first specific implementation method;

[0091] Then, the improved YOLOV5 algorithm is used to identify the target, and the improved part of the improved YOLOV5 algorithm is as follows:

[0092] (1) Up-sampling method replacement:

[0093] To ensure the real-time performance of the algorithm, YOLOV5s with a small network layer is selected as the basic model. The neck of YOLOV5 adopts a pyramid structure combining FPN and PAN for feature enhancement. Before inputting the pyramid structure, the image needs to be up-sampled three times to assist the fusion of the pyramid structure and the image. The nearest neighbor interpolation strategy is adopted for up-sampling, and the adjacent pixel points are copied to expand the number of pixel points.

[0094] The nearest neighbor interpolation strategy has small calculation amount, is simple and fast, but the expanded image will have a large degree of distortion, inlay, sawtooth and other phenomena that reduce the image quality. Convolution can down-sample an image, while deconvolution is the opposite operation of convolution, which can generate a picture from specific data to achieve the purpose of up-sampling. As convolution, deconvolution can also learn the network according to the characteristics of the data set in deep learning, obtain more suitable weight values for detecting targets, and better extract image features. Therefore, in the present application, deconvolution is used instead of the nearest neighbor interpolation strategy for up-sampling, which improves the detection index of the target detection algorithm, and the model size and algorithm frame rate are not increased.

[0095] (2) Data set cleaning:

[0096] The advantages and disadvantages of the data in the data set can affect the accuracy, average loss, overfitting speed of the deep learning model, and the deep learning model can more easily fit the noise. A large amount of Gaussian noise and impulse noise exists in the images in the data set, which can cause the classification algorithm accuracy to decrease significantly, so the less noise in the data set, the better the generalization ability of the model. After confidence learning on some currently public large data sets, it is detected that there are error labels in these data sets. The data set used for training the YOLOV5 algorithm is the COCO data set, so the data set needs to be cleaned to remove error labels and improve image quality.

[0097] In actual situation, the underwater operation target of underwater robot is relatively fixed, so only a few classification training targets are selected to identify the model. In order to reduce the experimental time cost, the data set is processed. First, random pictures are extracted from the COCO data set to reduce the data amount of the data set, and then only the labels and pictures containing the selected target classification are selected from the randomly extracted pictures to obtain the data set before cleaning. Artificial inspection of the labels of the data set before cleaning can find that there are some problems such as poor image quality, mismatch between label frame and classification, poor matching between label frame and actual target, and difficult to distinguish target still with label, etc. These problems are solved, the label frame is modified and improved, and the COCO cleaning data set used for training is obtained. Training the model with it can significantly improve the detection accuracy.

[0098] In addition to the improved part above, other processing procedures of the improved YOLOV5 algorithm are the same as the existing YOLOV5 algorithm.

[0099] The present application is aimed at complex underwater environment, using image processing algorithm and target recognition algorithm, which can realize image contrast increase and accurately locate the searched target. The system has wide application range and can be used for real-time processing, and is suitable for searching activities of people falling into river with turbulent water flow, low visibility and certain danger, so as to effectively improve the searching efficiency and eliminate the danger of manual rescue. The system is also suitable for observing the growth status of aquatic products in fishery breeding, so that the aquaculture personnel can achieve better benefits with clearer understanding of the aquatic products under smaller interference. In addition, the product also has good application prospect in assisting aquatic product fishing, ecological environment research and scientific exploration.

[0100] The present application also has other various embodiments, and those skilled in the art can make various corresponding changes and modifications according to the present application without departing from the spirit and essence of the present application, but these corresponding changes and modifications should belong to the protection scope of the claims attached to the present application.

Claims

1. An image processing method for assisting underwater robot operations, characterized in that, Includes the following steps: Step 1: Use the underwater robot's camera to capture video images, and record them as the original images; Step 2: The video image is processed using the improved DCP algorithm, which is the dark channel prior dehazing algorithm. The process of processing video images using the improved DCP algorithm includes the following steps: Obtain the image with the foggy dark channel and denote it as the original foggy dark channel image; For the original foggy dark channel image, the first step is to use a region segmentation method to distinguish between the light source area and the non-light source area. Then, the binarized image of the extracted non-light source area is multiplied by the dot product of the foggy dark channel image to obtain the foggy dark channel image after removing the light source. The foggy dark channel image after removing the light source is then fed into the atmospheric light component algorithm to obtain the atmospheric light component A1. The atmospheric light component A0 obtained by inputting the original foggy dark channel image into the atmospheric light component algorithm is the atmospheric light component intensity of the light source region. The final atmospheric light component A2 is obtained by weighted fusion of A0 and A1, and the specific mathematical expression is as follows: Where x represents the pixel position, y1 and y2 represent the pixel positions in the light source region and the non-light source region, respectively; α represents the fusion coefficient; Ar0(x) and Ar1(x) represent the light source region and the non-light source region, respectively; n0 and n1 represent the total number of pixels in the light source region and the total number of pixels in the non-light source region, respectively; the fusion coefficient α is as follows: TO 0_1 =In(1+A0) A 1_1 =In(1+A1) Where β is the correction coefficient; A 0_1 This refers to the atmospheric light composition of the light source region after processing the atmospheric light composition A0. 1_1 It is the atmospheric light component in the non-light source region after processing atmospheric light component A1; Transmittance was calculated using the atmospheric light composition A0 of the light source region: Where w represents a constant parameter for preserving depth of field, and I c (y) is the hazy image of the color channel; Ω(x) is the square region centered at x; The dehazed image is obtained using A2: Where t0 represents the lower bound of transmittance; I(x) represents a hazy image.

2. The image processing method for assisting underwater robot operations according to claim 1, characterized in that, The correction factor β is set to 0.

7.

3. The image processing method for assisting underwater robot operations according to claim 2, characterized in that, The constant parameter w for preserving depth of field is set to 0.

95.

4. The image processing method for assisting underwater robot operations according to claim 3, characterized in that, The lower limit of transmittance t0 is set to 0.

1.

5. An image processing method for assisting underwater robot operations according to any one of claims 1 to 4, characterized in that, The process of calculating transmittance using the atmospheric light composition A0 of the light source region is replaced by the following process: Where w represents a constant parameter for preserving depth of field, and I dark (x) represents an image with a foggy, dark channel.

6. An image processing method for assisting underwater robot operations according to any one of claims 1 to 4, characterized in that, In estimating atmospheric light composition, image resolution compression is first achieved by downsampling the hazy dark channel image by a factor of n, reducing the data size of the hazy dark channel image to 1 / n of its original size. 2 Then, the atmospheric light component is estimated by multiplying the pixel count of the image with the foggy dark channel and sorting all pixels.

7. An image processing method for assisting underwater robot operations according to any one of claims 1 to 4, characterized in that, In the process of processing video images using the improved DCP algorithm, the original image after grayscale is used as the guide map to perform guide filtering on the estimated transmittance. For cases where the guide filtering algorithm requires a large number of mean filters for image processing, a GPU parallel computing method is adopted to enable multiple mean filters to process the image simultaneously.

8. The image processing method for assisting underwater robot operations according to claim 7, characterized in that, The image used in the guided filtering process uses 64-bit floating-point data, while the images used in other processes use 32-bit floating-point data.

9. An image recognition method for assisting underwater robot operations, characterized in that, Includes the following steps: First, the video images captured by the underwater robot's camera are processed using the image processing method for assisting underwater robot operations as described in any one of claims 1 to 8; then, the improved YOLOv5 algorithm is used to identify the target. In the improved YOLOv5 algorithm, YOLOv5s is selected as the base model. A pyramid structure combining FPN and PAN is used to enhance features at the neck of YOLOv5. Before inputting the pyramid structure, the image needs to be upsampled three times, and deconvolution is used for upsampling during the three upsampling operations.

10. The image recognition method for assisting underwater robot operations according to claim 9, characterized in that, The improved YOLOV5 algorithm requires dataset cleaning before training. The improved YOLOV5 algorithm is trained using the cleaned dataset, thus obtaining the trained improved YOLOV5 algorithm. The process of dataset cleaning includes the following steps: First, images are randomly selected from the COCO dataset; Then, from the randomly selected images, only the labels and images containing the selected target category are selected to obtain the unwashed dataset. The labels of the unwashed dataset are manually checked and the label boxes are modified and improved to obtain the COCO cleaned dataset used to train and improve the YOLOv5 algorithm.

Images