An underwater visual SLAM system with adaptive image enhancement and its implementation method

Abstract

The present invention discloses an underwater visual SLAM system with adaptive image enhancement and an implementation method. The invention includes a low-light and turbid environment recognition module, a low-light image enhancement module, a turbid environment image enhancement module and an image enhancement processing module; the low-light and turbid environment recognition module is implemented by an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model; the low-light image enhancement module is implemented by an underwater low-light environment image enhancement algorithm based on CLAHE; the turbid environment image enhancement module is implemented by an underwater turbid environment image enhancement algorithm based on the dark channel prior principle; the image enhancement processing module is implemented by an image enhancement processing strategy based on serial-parallel fusion. The present invention can effectively restore underwater low-light and turbid images in a dynamic underwater environment, effectively solve the problem of reduced performance of underwater visual SLAM in underwater low-light and turbid environments, and improve the robustness and positioning accuracy of the robot's simultaneous positioning and mapping in a dynamic underwater environment.

Description

Technical Field

[0001] The invention belongs to the technical field of robots, and in particular relates to an underwater visual SLAM system with adaptive image enhancement and an implementation method thereof. Background Art

[0002] With the widespread application of underwater robots in the field of coral colony detection, the research on the simultaneous localization and mapping (SLAM) system of robots in underwater environments has attracted increasing attention. Unlike the controllable environment on the ground or indoors, the underwater environment is unstructured and has various noise interferences, which brings various difficulties and challenges to underwater visual SLAM. Seawater is a complex mixture of chemical elements, containing a large amount of suspended matter and organic mixtures. Due to light scattering and absorption, the contrast of the image is reduced. In addition, the inhomogeneity of seawater will cause greater image distortion. In low-light and turbid underwater environments, visual blur and feature changes will occur, which seriously affect the performance of the underwater visual SLAM system. Therefore, how to improve the performance of underwater visual SLAM in low-light and turbid underwater environments is an urgent problem that needs to be solved.

[0003] Huang et al. proposed an underwater visual SLAM based on the Retinex method of HSV color space guided filtering to enhance underwater images in order to solve the low resolution and color distortion problems of underwater images. Wu et al. proposed an underwater complex scene SLAM system based on image enhancement. The algorithm is based on the VINS-MONO algorithm framework, uses the FAST feature point extraction algorithm to improve the extraction speed, uses the inverse optical flow method to improve the accuracy of feature extraction, extracts several residual information at the edge of the back end and performs edge processing to improve the edge speed. Liu et al. proposed an underwater visual SLAM based on ESIHE image enhancement. The enhanced image can extract more ORB features and improve the attitude estimation accuracy of the SLAM method. The above research is committed to solving the impact of complex underwater environment on underwater visual SLAM system, and improves the feature extraction accuracy and attitude estimation accuracy by enhancing underwater images, thereby improving the robustness and positioning accuracy of underwater visual SLAM system. However, the current underwater visual SLAM system has not yet considered the dynamically changing underwater environment, and a single image enhancement algorithm cannot meet the real-time changes in the underwater environment. Therefore, how to improve the underwater visual SLAM performance when dealing with dynamically changing low-light and turbid underwater environments remains an unresolved problem. Summary of the invention

[0004] The present invention aims to solve the problem of performance degradation of underwater visual SLAM when facing dynamically changing low-light and turbid underwater environments. An underwater visual SLAM system with adaptive image enhancement and an implementation method are proposed. The system has stable robustness and positioning accuracy in underwater low-light and turbid environments, can accurately identify the current underwater environment, and adaptively enhance the collected underwater degraded images, effectively improving the visibility of underwater images, and can achieve accurate and robust simultaneous positioning and mapping of underwater robots even in the case of dynamic changes in the underwater environment.

[0005] The technical solution of the present invention is:

[0006] An adaptive image-enhanced underwater visual SLAM system, comprising a low-light and turbid environment recognition module, a low-light image enhancement module, a turbid environment image enhancement module, an image enhancement processing module, and a novel monocular visual SLAM based on an ORB-SLAM3 framework;

[0007] The low-light and turbid environment recognition module is implemented by an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model;

[0008] The low-light image enhancement module is implemented by an underwater low-light environment image enhancement algorithm based on CLAHE;

[0009] The turbid environment image enhancement module is implemented by an underwater turbid environment image enhancement algorithm based on the dark channel prior principle;

[0010] The image enhancement processing module is implemented by an image enhancement processing strategy based on serial-parallel fusion.

[0011] The present invention also provides a method for implementing an underwater visual SLAM system with adaptive image enhancement, which specifically comprises the following steps:

[0012] Step 1: Data collection and preprocessing;

[0013] Step 2: Recognition of underwater low-light environment images and underwater turbid environment images based on the AquaVisNet model;

[0014] If the recognition result is a low-light environment, then step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE; if the recognition result is a turbid environment, then step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle; if the recognition result is that both the low-light environment and the turbid environment exist, then step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE and step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle at the same time;

[0015] Step 3: Enhancement of underwater low-light environment images based on CLAHE;

[0016] Step 4: Enhancement of underwater turbid environment images based on dark channel prior principle.

[0017] The present invention introduces an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model and an underwater low-light and turbid environment image enhancement algorithm based on a serial-parallel fusion processing strategy in the tracking thread. The underwater low-light and turbid environment image enhancement algorithm based on a serial-parallel fusion processing strategy is combined with the SUB-CLAHE image enhancement algorithm and the DCP-UTI image enhancement algorithm proposed in the present invention to adaptively restore dynamically changing underwater degraded images, thereby maximizing the visibility of underwater images. In underwater low-light and turbid environments, the current underwater environment can be accurately identified, and the collected underwater degraded images can be adaptively enhanced, effectively improving the visibility of underwater images, thereby improving the performance of underwater visual SLAM.

[0018] The beneficial effects of the present invention are:

[0019] This paper proposes an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model, which can accurately identify the current environment in low-light and turbid underwater environments. At the same time, the algorithm trains the model by collecting the current underwater environment images in real time to ensure the robustness and accuracy of the algorithm in different underwater environments.

[0020] The present invention proposes an underwater low-light and turbid environment image enhancement algorithm based on a serial-parallel fusion processing strategy, which can effectively restore underwater low-light and turbid images in a dynamic underwater environment, improve image contrast, reduce image turbidity, enrich image texture information, and improve image feature point extraction and matching performance.

[0021] The present invention proposes an adaptive image-enhanced underwater visual SLAM for underwater low-light and turbid environments, which solves the problem of reduced performance of underwater visual SLAM in low-light and turbid environments and improves the robustness and positioning accuracy of the robot's simultaneous positioning and mapping in a dynamic underwater environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Schematic diagram of the experimental site;

[0023] Figure 2 Some images of the dataset; (a) normal underwater environment; (b) low-light underwater environment; (c) turbid underwater environment; (d) low-light and turbid underwater environment;

[0024] Figure 3 AquaVisNet model training diagram and training results;

[0025] Figure 4 SUB-CLAHE algorithm flow chart;

[0026] Figure 5 Image enhancement results in low-light environment; (a) original low-light image; (b) image enhanced using the CLAHE algorithm; (c) image enhanced using the SUB-CLAHE algorithm.

[0027] Figure 6 Image enhancement results in turbid environment; (a) original turbid image; (b) image enhanced using the DCP algorithm; (c) image enhanced using the DCP-UTI algorithm.

[0028] Figure 7 Flowchart of image enhancement processing strategy based on serial and parallel fusion;

[0029] Figure 8 AIE-ORB-SLAM system structure diagram;

[0030] Fig. 9 SLAM operation result data; (a) normal environment; (b) low-light environment; (c) turbid environment; (d) low-light and turbid environment.

[0031] Fig.10 Estimated trajectory maps of ORB-SLAM3 and AIE-ORB-SLAM; (a) normal environment trajectory map; (b) low-light environment trajectory map; (c) turbid environment trajectory map; (d) low-light and turbid environment trajectory map. DETAILED DESCRIPTION

[0032] The present invention is further described below with reference to the accompanying drawings.

[0033] An adaptive image-enhanced underwater visual SLAM system, comprising a low-light and turbid environment recognition module, a low-light image enhancement module, a turbid environment image enhancement module, an image enhancement processing module, and a novel monocular visual SLAM based on an ORB-SLAM3 framework;

[0034] The low-light and turbid environment recognition module is realized by an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model; the low-light image enhancement module is realized by an underwater low-light environment image enhancement algorithm based on CLAHE; the turbid environment image enhancement module is realized by an underwater turbid environment image enhancement algorithm based on the dark channel prior principle; the image enhancement processing module is realized by an image enhancement processing strategy based on serial-parallel fusion.

[0035] Furthermore, the present invention provides a method for implementing an underwater visual SLAM system with adaptive image enhancement, and the specific implementation steps are as follows:

[0036] Step 1: Data collection and preprocessing;

[0037] Step 2: Recognition of underwater low-light and turbid environment images based on the AquaVisNet model;

[0038] Step 3: Enhancement of underwater low-light images based on CLAHE;

[0039] Step 4: Enhancement of underwater turbid environment images based on dark channel prior principle;

[0040] like Figure 1 As shown, the data collection and preprocessing described in step 1 of the present invention are specifically as follows:

[0041] The present invention has built a water tank experimental system simulating an ocean environment, such as Figure 1 As shown. The experimental site consists of a pool, a Blue ROV2, and an artificial coral colony. The present invention randomly arranges different varieties of artificial corals to form an artificial coral colony to simulate the state of coral colonies in the ocean. In order to evaluate the effect of the AIE-ORB-SLAM system, the present invention controls the ROV to shoot underwater environment data sets of normal, low light, turbidity, and low light and turbidity at the same time, as shown in the figure. Figure 2 As shown in Figure 2, the same dataset is imported into the AIE-ORB-SLAM system and the ORB-SLAM3 system to compare the robustness and positioning accuracy of the two SLAM systems under the same dataset.

[0042] Further, such as Figure 3 As shown, the recognition of underwater low-light and turbid environment based on the AquaVisNet model described in step 2 of the present invention is as follows:

[0043] The AquaVisNet model has seven layers, including: 2 convolutional layers, 2 pooling layers, 1 Flatten layer, and 2 fully connected layers. The first convolutional layer has 32 filters of size 3×3 and ReLU activation function. This layer accepts images of size 128×128 pixels with 3 channels (red, green, and blue). The second pooling layer uses a 2×2 pooling window to downsample the output of the previous layer, retaining the most significant features and reducing the output dimension. The third convolutional layer has 64 filters of size 3×3 and ReLU activation function to extract more abstract features of the image. The fourth pooling layer uses a 2×2 pooling window to further reduce the input tensor to accommodate the subsequent fully connected layer. The fifth Flatten layer expands the feature map from a multi-dimensional shape into a one-dimensional feature vector for use by the fully connected layer. The sixth fully connected layer has 64 neuron nodes and uses the ReLU activation function. The last fully connected layer includes 4 neurons, corresponding to the number of output categories of the model. The softmax activation function is used to convert the output of the model into probability values ​​of the 4 output categories for identifying different underwater environments.

[0044] If the current environment recognition result is a low-light environment, step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE; if it is a turbid environment, step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle; if it is a low-light environment and a turbid environment, step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE and step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle at the same time.

[0045] 8498 underwater coral images taken in a pool experimental system simulating the marine environment were used as a dataset. After removing low-quality images, 4000 samples were randomly selected from the dataset as the training set of the AquaVisNet model, and 1712 samples in the dataset were selected as the validation set. During the model training process, the number of training set iterations was set to 10, and the batch size was set to 32 to improve the performance of the network model.

[0046] After the AquaVisNet model training was completed, the accuracy of the AquaVisNet model for the training set reached 97.23%, and the accuracy of the validation set reached 96.54%; the loss value of the training set was 0.15, and the loss value of the validation set was 0.17. After continuous debugging, the AquaVisNet model achieved an environmental recognition accuracy of 99.58% in underwater low-light and turbid environments.

[0047] Further, such as Figure 4 As shown, the image enhancement in an underwater low-light environment based on CLAHE (SUB-CLAHE) described in step 3 of the present invention mainly includes the following steps:

[0048] 3-1. Separate the underwater low-light image into three channels: B, G, and R. The image of each channel is divided into original local image blocks of the same size so that each original local image block can be processed independently.

[0049] 3-2. For each original local image block, calculate its grayscale histogram ,in Indicates the gray levels that appear.

[0050] 3-3. Calculate the average number of pixels assigned to each gray level :

[0051]

[0052] in, Indicates the number of pixels in the horizontal direction of the local image block; Indicates the number of pixels in the vertical direction; Indicates the number of gray levels in a local image block.

[0053] 3-4. Based on the grayscale histogram Set the grayscale histogram threshold based on the distribution of In low-light underwater environments, the threshold For dynamic adaptation to lighting changes and lack of contrast;

[0054]

[0055] in, is the grayscale histogram clipping limit coefficient of underwater low-light images.

[0056] 3-5. According to the limited threshold corresponding to each original local image block , the grayscale histogram of the original local image block Exceeding the limit threshold The cut part is cut and evenly distributed to each gray level. After a round of distribution, the gray level histogram of the local image block will rise to a certain height, such as Figure 2 As shown, the limit threshold is exceeded again. , so repeat the above clipping operation until the grayscale histogram no longer exceeds the limited threshold .

[0057] 3-6. Based on the grayscale histogram after clipping, perform histogram equalization on each original local image block. This step aims to balance the distribution of pixel intensity within each original local image block, reduce the adverse effects of uneven illumination, and improve the overall image quality.

[0058] 3-7. According to the location of the original local image block, each pixel of the original local image block is processed by bilinear interpolation to reconstruct the pixel grayscale value. The enhanced B, G and R channel images are re-merged into an enhanced color image. This step ensures a smooth transition between local equalization areas, making the enhanced image more visually coherent and information-rich.

[0059] An image is randomly selected from the dataset and processed using the CLAHE image enhancement algorithm and the SUB-CLAHE image enhancement algorithm. The processing results are shown in the figure. Figure 5 As shown in Table 1, a quantitative comparison of the original underwater low-light image, the underwater image enhanced by the CLAHE algorithm, and the underwater image enhanced by the SUB-CLAHE algorithm is shown. After the SUB-CLAHE image enhancement algorithm is used, the UCIQE index of the enhanced image is improved by 22.16% and the UIQM index is improved by 22.44% compared with the original image; compared with the image enhanced by the CLAHE algorithm, the UCIQE index is improved by 54.88% and the UIQM index is improved by 92.33%.

[0060] Table 1 Comparison of no-reference quality evaluation of underwater low-light images

[0061] method UCIQE UIQM Raw underwater low-light images 0.4221 1.7651 CLAHE 0.2447 0.1764 SUB-CLAHE 0.5423 2.2758

[0062] Furthermore, the underwater turbid environment image enhancement (DCP-UTI) based on the dark channel prior principle described in step 4 of the present invention is specifically implemented as follows:

[0063] The underwater turbidity image model is defined as:

[0064]

[0065] in: This is the original underwater turbidity image; It is a clear image enhanced by the DCP-UTI algorithm; It is the background light component in the underwater environment; It stands for underwater transmittance, which is the degree to which light can pass through the underwater environment.

[0066] For the underwater turbidity image model, it is normalized, that is, both sides are divided by the underwater environment background light value of each channel, and the result is:

[0067]

[0068] in, Indicates B, G and R channels. Assuming the background light value of the underwater environment is a known quantity, The underwater transmittance is a constant and is defined as , and perform two minimization operations on both sides of the above equation, we can get:

[0069]

[0070] For clear images The underwater dark channel tends to 0, that is: have:

[0071]

[0072] in, Take the minimum value of the RGB value in a window as the corresponding dark channel pixel value. represents the dark channel pixel value, Represents the pixel value of a channel of the image; Indicates that a center in the image is located at A window of points, centered at point; express The point is inside the window; It means taking the minimum pixel value at a certain point y in the window; c refers to any RGB channel. Refers to taking the minimum value from the RBG channel.

[0073] roll out:

[0074]

[0075] Substituting back into equation (5), we get:

[0076]

[0077] In order to prevent the underwater image from being too clear and the restored scene from being unnatural, the underwater turbidity limit parameter is introduced. , we can get:

[0078]

[0079] The above derivation process assumes that the background light in the underwater environment is It is known that in practice, the background light The underwater dark channel map is used to obtain the original underwater turbid image. The specific steps are as follows: first obtain the underwater dark channel map and extract the brightest top 0.1% pixels according to the brightness; Find the value of the point with the highest brightness at the corresponding position as the background light in the underwater environment .

[0080] Considering the underwater transmittance When the value is very small, it will cause A large value will cause the entire image to transition to white, so a threshold is set. ,when < season = Substituting the underwater transmittance and the background light value in the underwater environment into the formula, the final restoration formula of the underwater turbid image is as follows:

[0081]

[0082] An image is randomly selected from the dataset and processed using the DCP image enhancement algorithm and the DCP-UTI image enhancement algorithm. The processing results are shown in the figure below. Figure 6 As shown in Table 2, the quantitative comparison of the original underwater turbidity image, the underwater image enhanced by the DCP algorithm, and the underwater image enhanced by the DCP-UTI algorithm is shown, and the best result is marked in bold. From the experimental results, it can be observed that after the DCP-UTI image enhancement algorithm is processed, the UCIQE index of the enhanced image is improved by 24.06% and the UIQM index is improved by 24.56% compared with the original image; compared with the image enhanced by the DCP algorithm, the UCIQE index is improved by 12.99% and the UIQM index is improved by 18.32%.

[0083] Table 2 Comparison of no-reference quality evaluation of underwater turbidity images

[0084] method UCIQE UIQM Original underwater turbidity image 0.4384 1.9321 DCP 0.5023 2.0919 DCP-UTI 0.5773 2.5612

[0085] like Figure 7 As shown, the present invention adopts an image enhancement processing strategy based on serial-parallel fusion. After the ROV collects the image of the current environment, it transmits it to the system based on optical property perception and underwater image quality assessment. The system automatically selects appropriate image enhancement parameters and image enhancement processing methods based on the current underwater environment recognition results and image quality assessment results. In underwater low-light environments, the system can adaptively adjust exposure compensation and contrast parameters to improve visibility. In underwater turbid environments, the system can adaptively adjust denoising and clarity parameters. Finally, the images that have undergone the two image enhancement processes are weighted fused to obtain an underwater restored image that is visualized to the greatest extent.

[0086] like Figure 8As shown, the present invention proposes an adaptive image enhanced underwater visual SLAM system and implementation method (AIE-ORB-SLAM). The AIE-ORB-SLAM system consists of three parallel threads (tracking thread, local underwater mapping thread, closed-loop detection and underwater map fusion thread) and a multi-underwater map management system. In the tracking thread, an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model and an underwater low-light and turbid environment image enhancement algorithm based on a serial-parallel fusion processing strategy are introduced to solve the image degradation problem in low-light and turbid underwater environments, and improve the robustness and positioning accuracy of underwater robots in simultaneous positioning and mapping in dynamic underwater environments.

[0087] Robustness evaluation

[0088] During the operation of the underwater visual SLAM system, the main parameters such as initialization, tracking loss, key frames, point clouds, etc. are counted, such as Fig. 9 shown.

[0089] In the normal environment dataset, ORB-SLAM3 and AIE-ORB-SLAM have similar performances in terms of main parameters such as initialization, tracking loss, keyframes, and point clouds, with no significant differences.

[0090] In the low-light environment dataset, the initialization time of the AIE-ORB-SLAM system is improved by 11.32%, the duration is improved by 19.08%, the number of key frames is improved by 37.33%, and the number of point clouds is improved by 31.66% compared with the ORB-SLAM3 system.

[0091] In the turbid environment dataset, the initialization time of the AIE-ORB-SLAM system is improved by 6.56%, the duration is improved by 53.85%, the number of key frames is improved by 55.26%, and the number of point clouds is improved by 68.68% compared with the ORB-SLAM3 system.

[0092] In the low-light and turbid environment dataset, the initialization time of the AIE-ORB-SLAM system is improved by 12.89%, the duration is improved by 59.34%, the number of key frames is improved by 61.72%, and the number of point clouds is improved by 69.44% compared with the ORB-SLAM3 system.

[0093] For underwater visual SLAM systems, faster initialization, longer tracking time, and more key frames and point clouds represent better robustness of the underwater visual SLAM system. Therefore, from the above experimental results, it can be concluded that in underwater low-light and turbid environments, the AIE-ORB-SLAM system exhibits better robustness than the ORB-SLAM3 system.

[0094] Accuracy evaluation

[0095] The present invention adopts absolute trajectory error (ATE) to evaluate the accuracy of underwater visual SLAM system. Fig.10 The ORB-SLAM3 estimated trajectory, AIE-ORB-SLAM estimated trajectory and true trajectory are shown. The experimental results show that in four different underwater environments, the ORB-SLAM3 system is prone to tracking loss, while the AIE-ORB-SLAM system can still generate a stable estimated trajectory that is close to the true trajectory.

[0096] To further verify the accuracy performance of AIE-ORB-SLAM, we compared the absolute trajectory error (ATE) between the estimated trajectory and the actual trajectory of the ORB-SLAM3 and AIE-ORB-SLAM systems, as shown in Table 3. In this dataset, the positioning accuracy of the AIE-ORB-SLAM system is improved by 50.17% in normal environments, 84.34% in low-light environments, 20.29% in turbid environments, and 35.83% in low-light and turbid environments compared to the ORB-SLAM3 system. It can be seen that the positioning accuracy of AIE-ORB-SLAM is better than that of ORB-SLAM3 in underwater low-light and turbid environments.

[0097] Table 3 Average ATE and RMS ATE of underwater visual SLAM system on the dataset

[0098]

[0099] Comprehensive experimental results show that the AIE-ORB-SLAM system proposed in this invention has better robustness and higher positioning accuracy than the ORB-SLAM3 system in underwater low-light and turbid environments, and can achieve accurate underwater robot positioning and map construction.

Claims

1. A method for implementing an underwater visual SLAM system with adaptive image enhancement, characterized in that: The specific implementation steps are as follows: Step 1: Data collection and preprocessing; Step 2: Recognition of underwater low-light environment images and underwater turbid environment images based on the AquaVisNet model; If the recognition result is a low-light environment, then step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE; if the recognition result is a turbid environment, then step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle; if the recognition result is that both the low-light environment and the turbid environment exist, then step 3 is performed to enhance the image in the underwater low-light environment based on CLAHE and step 4 is performed to enhance the image in the underwater turbid environment based on the dark channel prior principle at the same time; Step 3: Enhancement of underwater low-light environment images based on CLAHE; Step 4: Enhancement of underwater turbid environment images based on dark channel prior principle; Step 3 contains the following steps: 3-1. Separate the underwater low-light environment image into three channels: B, G and R; divide the image of each channel into original local image blocks of the same size so that each original local image block can be processed independently; 3-2. Calculate the grayscale histogram of each original local image block ,in Indicates the gray level that appears; 3-3. Calculate the average number of pixels assigned to each gray level : ； in, Indicates the number of pixels in the horizontal direction of the original local image block; Indicates the number of pixels in the vertical direction; Represents the number of gray levels in the original local image block; 3-4. Based on the grayscale histogram The distribution of grayscale histogram is set to the limit threshold ; In low-light underwater environments, limit the threshold For dynamic adaptation to lighting changes and lack of contrast; ； in, is the shear limit coefficient of the grayscale histogram corresponding to the underwater low-light environment image; 3-5. According to the limited threshold corresponding to each original local image block , the grayscale histogram of the original local image block Exceeding the limit threshold The cut part is cut and evenly distributed to each gray level; after one round of distribution, the gray histogram of the original local image block will rise to a certain height, and then it will exceed the limited threshold again. , so repeat the above clipping operation until the grayscale histogram no longer exceeds the limited threshold ; 3-6. Based on the grayscale histogram after cutting, perform histogram equalization processing on each original local image block; 3-7. According to the position of the original local image block, each pixel of the original local image block is processed by bilinear interpolation to reconstruct the pixel gray value; the enhanced B, G and R channel images are re-merged into an enhanced color image; Step 4 is implemented as follows: 4-1. Define the underwater turbidity image model as: ； in: This is the original underwater turbidity image; It is a clear image enhanced by the DCP-UTI algorithm; It is the background light value component in the underwater environment; represents underwater transmittance, which is the degree to which light can pass through the underwater environment; 4-2. Normalize the underwater turbidity image model and divide both sides by the underwater environment background light value of each channel to obtain: ； in, Represents B, G, and R channels; assuming the background light value of the underwater environment is a known quantity, The underwater transmittance is a constant and is defined as , perform two minimization operations on both sides of the above equation, and we get: ； Due to the clear image The underwater dark channel tends to 0, that is: ,therefore: ； in, Take the minimum value of the RGB value in a window as the corresponding dark channel pixel value; where, represents the dark channel pixel value, Represents the pixel value of a channel of the image; Indicates that a center in the image is located at A window of points, centered at point; express The point is inside the window; Indicates taking a point in this window The pixel value is the minimum; c refers to any RGB channel; Refers to taking the minimum value from the RBG channel; roll out: ； Substituting back into equation (5), we get: ； In order to prevent the underwater image from being too clear and the restored scene from being unnatural, the underwater turbidity limit parameter is introduced. ,have to: ； Considering the underwater transmittance When the value is very small, it will cause A large value will cause the entire image to transition to white, so a threshold is set. ,when season ; Substitute the underwater transmittance and the background light value in the underwater environment obtained above into the formula, and finally get the restoration formula of underwater turbid image as follows: 。 2. The method for implementing an underwater visual SLAM system with adaptive image enhancement as claimed in claim 1, characterized in that: The data collection and preprocessing described in step 1 are as follows: A water pool experimental system simulating the marine environment was built, which consisted of a water pool, a Blue ROV2, and an artificial coral colony. Different varieties of artificial corals were randomly arranged to form an artificial coral colony to simulate the state of the coral colony in the ocean. The underwater environment dataset was collected by driving the Blue ROV2. The underwater environment dataset included normal images, underwater low-light environment images, underwater turbid environment images, and underwater low-light and turbid environment images.

3. The method for implementing an underwater visual SLAM system with adaptive image enhancement as claimed in claim 1, characterized in that: Step 2 is as follows: The construction of the AquaVisNet model, which has seven layers, including: 2 convolutional layers, 2 pooling layers, 1 Flatten layer and 2 fully connected layers; The first convolutional layer has 32 filters of size 3×3 and ReLU activation function. This layer receives images of size 128×128 pixels with 3 channels. The second pooling layer uses a 2×2 pooling window to downsample the output of the previous layer; The third convolutional layer has 64 filters of size 3×3 and ReLU activation function; The fourth pooling layer uses a 2×2 pooling window to further reduce the input tensor; The fifth layer, Flatten, expands the feature map from a multi-dimensional shape into a one-dimensional feature vector for use by the fully connected layer; The sixth fully connected layer has 64 neuron nodes and uses the ReLU activation function; The last fully connected layer includes 4 neurons, corresponding to the number of output categories of the model. The softmax activation function is used to convert the output of the model into probability values ​​of the 4 output categories for identifying different underwater environments.

4. The method for implementing an underwater visual SLAM system with adaptive image enhancement as claimed in claim 1, characterized in that: Background light value The underwater dark channel map is used to obtain the original underwater turbid image. The specific steps are as follows: first obtain the underwater dark channel map and extract the brightest top 0.1% pixels according to the brightness; Find the value of the point with the highest brightness at the corresponding position as the background light value in the underwater environment .

5. The method for implementing an adaptive image-enhanced underwater visual SLAM system as claimed in claim 1, characterized in that: The system includes a low-light and turbid environment recognition module, a low-light image enhancement module, a turbid environment image enhancement module, and an image enhancement processing module; The low-light and turbid environment recognition module is realized by an underwater low-light and turbid environment recognition algorithm based on the AquaVisNet model; the low-light image enhancement module is realized by an underwater low-light environment image enhancement algorithm based on CLAHE; the turbid environment image enhancement module is realized by an underwater turbid environment image enhancement algorithm based on the dark channel prior principle; the image enhancement processing module is realized by an image enhancement processing strategy based on serial-parallel fusion.

Images