Image processing system and method suitable for turbid water area

Abstract

The invention discloses an image processing system and method suitable for a turbid water area, and belongs to the technical field of image processing. The system comprises an underwater camera component for collecting original image data of a turbid water area in real time and an embedded image processing unit. The embedded image processing unit comprises a preprocessing module used for carrying out noise suppression and image calibration processing on original image data, and a physical parameter estimation module used for estimating environment background light intensity of a current water area, a water body attenuation coefficient and a distance map of an object in a scene and calculating initial transmissivity. The scattering correction module is used for removing back scattering and preliminarily compensating a color channel based on a physical imaging model in combination with physical parameters, and the deep learning enhancement module is used for improving image local contrast, adjusting brightness and finely correcting colors by adopting a lightweight deep convolutional neural network. According to the image processing system and method suitable for the turbid water area, the image definition and color of the turbid water area can be improved, the real-time performance is high, and the system and method are suitable for different water area environments.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of image processing, and particularly relates to an image processing system and method suitable for turbid water. BACKGROUND

[0002] In the process of underwater imaging, due to the absorption and scattering of light in water, problems such as color distortion and reduced visibility often occur, especially in turbid water, a large number of suspended particles cause strong light scattering, resulting in a significant decrease in image contrast and clarity. Existing underwater image enhancement techniques mainly fall into three categories: Firstly, physical model methods such as histogram equalization, white balance, and Retinex algorithms are simple to implement and effective for lightly degraded images, but they do not consider the imaging physical process, making it difficult to eliminate the turbidity caused by scattering in complex underwater environments, and they are prone to over-saturation or color imbalance defects; Secondly, methods based on optical transmission models treat underwater imaging as a special case of atmospheric scattering models, and infer the image by establishing a model, such as the Jaffe-McGlamery imaging model, which requires estimating the transmittance of each pixel and the global background light intensity to restore the clear image. However, it is difficult to accurately estimate the water environment parameters, and the pixel-by-pixel iterative calculation has a heavy load and poor real-time performance, which is not conducive to on-site deployment of hardware; Thirdly, learning methods use convolutional neural networks to learn the mapping relationship from degraded to clear data, which performs well on ideal data sets, but relies heavily on large-scale high-quality training data, it is difficult to obtain underwater data sets covering various water environments and real clear "ground truth", the model generalization ability is poor, and most deep networks have large computational load, making it difficult to run in real time on embedded hardware.

[0003] In recent years, hybrid methods that combine physical models and deep learning have emerged, which to some extent alleviate the dependence on real data for pure data-driven methods, but still face challenges such as accurate parameter estimation and complex model integration. In summary, the existing technology has problems such as insufficient adaptability to different water environments, incomplete recovery of low contrast caused by scattering, high dependence on real parameters or large data requirements, and difficulty in real-time processing in turbid water image enhancement. SUMMARY

[0004] In view of the deficiencies in the related art, the present application aims to provide an image processing system and method suitable for turbid water to solve the problems raised in the background art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions: An image processing system suitable for turbid water, comprising an underwater camera assembly and an embedded image processing unit. The underwater camera assembly is used to collect raw image data of turbid water in real time and transmit the collected raw image data to an embedded image processing unit; The embedded image processing unit comprises a preprocessing module, a physical parameter estimation module, a scattering correction module and a deep learning enhancement module connected in sequence; The preprocessing module is used for noise suppression and image calibration processing of the raw image data; The physical parameter estimation module is used for estimating the environmental background light intensity, water attenuation coefficient and object distance map in the scene of the current water area, and calculating the initial transmittance; The scattering correction module is used for removing backward scattering and performing preliminary compensation processing on the image data processed by the preprocessing module based on the physical imaging model and the parameters output by the physical parameter estimation module; The deep learning enhancement module is used for performing local contrast enhancement, brightness adjustment and color fine correction processing on the image data processed by the scattering correction module by using a lightweight deep convolutional neural network, and outputting an enhanced clear image.

[0006] In some embodiments, the physical parameter estimation module includes a background light estimation submodule, a water attenuation coefficient estimation submodule, a depth distance estimation submodule and an initial transmittance calculation submodule; The background light estimation submodule estimates and determines the environmental background light intensity by analyzing the histogram and color information of the image data and selecting the area with the lightest color or the highest brightness in the image; The water attenuation coefficient estimation submodule determines the water attenuation coefficient of each color channel by evaluating the turbidity of the water body; The depth distance estimation submodule estimates the object distance map in the scene by using a pre-trained monocular depth neural network, binocular camera ranging and fusion of external ranging information; The initial transmittance calculation submodule calculates the initial transmittance based on the water attenuation coefficient and the object distance map:

[0007] wherein, is the initial transmittance, is a natural constant, is the average value of the water attenuation coefficient, is the object distance map in the scene.

[0008] In some embodiments, the scattering correction module includes a backward scattering removal submodule and a color channel preliminary compensation submodule; The backward scattering removal submodule calculates the image after removing backward scattering based on the physical imaging model and combining the environmental background light intensity and the initial transmittance:

[0009] in, To remove the backscattered image, For the original image, For ambient background light intensity, The initial transmittance, To prevent division by zero of small constants; The color channel preliminary compensation submodule determines the gain of each color channel according to the attenuation law of each wavelength of light in water and the degree of attenuation of the corresponding wavelength of light. It then performs channel-by-channel gain compensation on the image after removing backscattering. The color channel with a higher degree of attenuation has a higher gain setting.

[0010] In some embodiments, the lightweight deep convolutional neural network is an encoder-decoder model. The encoder uses a depthwise separable convolutional or group convolutional structure to extract multi-scale spatial features of the image. The decoder uses a bilinear upsampling, fast convolutional kernel, and lightweight residual block structure to improve local contrast, adjust brightness, and finely correct color. The encoder and decoder transfer multi-scale spatial features through skip connections. A conditional normalization module is set in the network feature fusion layer of the encoder and decoder. The conditional normalization module introduces the initial transmittance and ambient background light intensity output by the physical parameter estimation module as conditional constraints to dynamically adjust the normalization parameters and activation function coefficients of the lightweight deep convolutional neural network.

[0011] In some embodiments, the embedded image processing unit further includes a frame rate monitoring module, which monitors the frame rate of the image enhancement processing in real time. When the frame rate is lower than a preset threshold, it automatically adjusts the inference rate of the lightweight deep convolutional neural network to keep the processing latency within a preset range.

[0012] In some embodiments, the embedded image processing unit further includes a display unit module, which is connected to the deep learning enhancement module and is used to receive and display the enhanced clear image output by the deep learning enhancement module in real time. The display unit module supports adaptation to display devices with different resolutions according to the actual application scenario.

[0013] An image processing method suitable for turbid waters, employing the aforementioned image processing system for turbid waters, includes the following steps: S1. The underwater camera component collects raw image data of turbid water in real time and transmits the raw image data to the embedded image processing unit. S2. The preprocessing module is used to perform noise suppression and image calibration on the original image data to obtain the preprocessed image. S3, estimating the ambient background light intensity, water attenuation coefficient and object distance map in the scene of the current water area by the physical parameter estimation module, and calculating the initial transmittance; S4, based on the physical imaging model, combining the ambient background light intensity and the initial transmittance obtained in step S3, using the scatter correction module to perform backward scattering removal and color channel preliminary compensation processing on the preprocessed image data, to obtain the scatter corrected image; S5, using a lightweight deep convolutional neural network, the scatter corrected image is processed by the deep learning enhancement module to enhance the local contrast, adjust the brightness and fine correct the color, and output the enhanced clear image.

[0014] In some embodiments, in step S3, the specific process of estimating the ambient background light intensity is: analyzing the histogram and color information of the original image data, counting the brightness values of each region in the image, selecting the region with the highest brightness value, and taking the average light intensity of the region as the ambient background light intensity; the specific process of calculating the initial transmittance is: based on the obtained water attenuation coefficient of the current water area and the object distance in the scene, the initial transmittance is calculated by using the formula , wherein, is the initial transmittance, is a natural constant, is the average value of the water attenuation coefficient, is the object distance map in the scene.

[0015] In some embodiments, in step S4, the specific process of backward scattering removal is: based on the physical imaging model, combining the ambient background light intensity and the initial transmittance, using the formula to calculate the image after removing the backward scattering, the formula is: , wherein, is the image after removing the backward scattering, is the original image, is the ambient background light intensity, is the initial transmittance, is a small constant to prevent division by zero; the specific process of color channel preliminary compensation is: according to the attenuation law of each wavelength of light in water, the gain of each color channel is determined according to the attenuation degree of the corresponding wavelength of light, and the gain compensation is performed on the image after removing the backward scattering, wherein the gain of the color channel corresponding to the wavelength with higher attenuation degree is set to be higher.

[0016] In some embodiments, in step S5, the training process of the lightweight deep convolutional neural network is trained until convergence using a joint optimization strategy of physical consistency loss and multi-scale perception loss; the physical consistency loss is used to constrain the output image of the lightweight deep convolutional neural network to remain consistent with the original input image in the physical space after degradation through the physical imaging model, and the multi-scale perception loss is used to optimize the pixel-level difference between the output image of the lightweight deep convolutional neural network and the target clear image.

[0017] Compared with the prior art, the present application has the following advantages: 1、The image processing system and method suitable for turbid water area provided by the present application break through the limitations of traditional single method, fuse underwater optical imaging physical priori and neural network learning, use physical model to accurately depict the scattering degradation process and embed network input and training, ensure that the enhanced result has physical credibility, avoid the unrealistic effects easily produced by pure learning method; at the same time, the neural network makes up for the defects of physical method in accurately modeling complex scenes, the advantages of the two are complementary, compared with the application of physical model or deep learning alone, the turbid water image enhancement performance can be more outstanding.

[0018] 2、The image processing system and method suitable for turbid water area provided by the present application, aiming at the real-time application requirements underwater, through algorithm and implementation optimization, through module splitting and parallel pipeline processing, the image acquisition and result output are efficiently connected; a lightweight neural network model is introduced, the network parameter quantity is small, combined with bilinear grid upsampling and other acceleration strategies, the operation efficiency and high resolution processing capability are considered, the embedded device is adapted, and the method can be directly deployed on underwater robots or camera terminals, online rapid enhancement is realized, and the disadvantages of offline processing of large deep models are overcome.

[0019] 3、The image processing system and method suitable for turbid water area provided by the present application have good adaptability to different turbidity and depth of water environment, rely on the physical parameter estimation module, can dynamically adjust the transmittance and gain compensation strategy according to the current water quality, and ideal enhancement effect is obtained in shallow sea, deep sea, fresh water, sea water and other scenes, which breaks through the specific limitations of existing methods in water area; and because the network is trained with physical consistency constraint, the network can adapt to the change of environment in the range of training data, and is not easy to fail under unknown water body conditions, and the dependence on specific data set is greatly reduced.

[0020] 4、The image processing system and method suitable for turbid water area provided by the present application solve the defects of the prior art that only focuses on a single problem, and through multi-module cooperation, multiple target processing such as despeckling, brightening, color correction, and sharpening is simultaneously realized: the physical module first improves the global visibility and contrast, the deep learning module further optimizes the detail quality and color authenticity, and the output image is superior to a single means in terms of definition, contrast, and color; the wideband attenuation prior (based on multi-channel transmission characteristics) is combined to guide network training, the complex light recovery effect is improved, and even in an extreme turbid environment, a natural and clear enhanced image can be output. BRIEF DESCRIPTION OF DRAWINGS

[0021] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their description serve to explain the present application. They do not constitute an improper limitation on the present application. In the drawings: Figure 1 A method flowchart of one embodiment of the image processing system and method suitable for turbid water area of the present application; Figure 2 A function module and workflow schematic diagram of one embodiment of the image processing system and method suitable for turbid water area of the present application; Figure 3 An algorithm flowchart of one embodiment of the image processing system and method suitable for turbid water area of the present application; Figure 4 A lightweight deep convolutional neural network structure diagram of one embodiment of the image processing system and method suitable for turbid water area of the present application; Figure 5 A lightweight deep convolutional neural network parameter schematic diagram of one embodiment of the image processing system and method suitable for turbid water area of the present application. DETAILED DESCRIPTION

[0022] The technical solutions in the embodiments will be described in detail below with reference to the drawings of the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments of the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0023] In the description of the present application, it needs to be understood that the terms "center", "transverse", "longitudinal", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship shown in the drawings, which are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation of the present application.

[0024] In the description of the present application, it needs to be understood that unless otherwise explicitly specified and limited, the terms "mounting", "connecting", "connecting" should be understood broadly, for example, it can be fixedly connected, or it can be detachably connected, or integrally connected; it can be directly connected, or indirectly connected through an intermediate medium, or it can be the communication inside two elements. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0025] Embodiment 1: Referring to the accompanying drawings Figures 1 to 5 An illustrative embodiment of the image processing system suitable for turbid water area is given, which includes an underwater camera assembly and an embedded image processing unit. The underwater camera assembly is used to collect raw image data of turbid water area in real time, and transmit the collected raw image data to the embedded image processing unit.

[0026] The image processing system suitable for turbid water area can obtain clear enhanced images in an environment where the water quality is turbid and ordinary cameras can hardly see clearly. The underwater camera assembly can select a conventional underwater camera, and a fill light can be used if necessary. The embedded image processing unit is deployed with software algorithms, which can be realized by an on-board GPU of an underwater robot or a customized hardware circuit.

[0027] Referring to the accompanying drawings Figure 2 The functional modules and working process of the image processing system of the embodiment are shown in the accompanying drawings. The embedded image processing unit includes a preprocessing module, a physical parameter estimation module, a scattering correction module and a deep learning enhancement module connected in sequence.

[0028] The preprocessing module is used for noise suppression and image calibration of the raw image data. Specifically, the raw turbid water area image data is obtained from the underwater camera assembly, and after preprocessing, the image taken by the underwater camera assembly is transmitted to the subsequent processing module. For video stream, the system processes frame by frame continuously.

[0029] The physical parameter estimation module is configured to estimate the ambient background light intensity, the water attenuation coefficient and the object distance map in the scene of the current water area, and calculate the initial transmittance. Specifically, the physical parameter estimation module comprises a background light estimation submodule, a water attenuation coefficient estimation submodule, a depth distance estimation submodule and an initial transmittance calculation submodule. The input image is quickly analyzed by each submodule to estimate the preliminary physical imaging parameters of the water body.

[0030] Firstly, the background light estimation submodule selects the area with the lightest color or the highest brightness in the image by analyzing the histogram and color information of the image data, and estimates and determines the ambient background light intensity . Then, the water attenuation coefficient estimation submodule estimates the distribution of the atmospheric transmittance of each color channel and the range of the water attenuation coefficient value by evaluating the turbidity of the water body. In this embodiment, an auxiliary depth / distance estimation means is introduced: the depth distance estimation submodule estimates the object distance map in the scene by using a pre-trained monocular depth neural network, binocular camera ranging and fusing external ranging information. The initial transmittance calculation submodule calculates the initial transmittance based on the water attenuation coefficient and the object distance map .

[0031] wherein, is the initial transmittance, is a natural constant, is the average value of the water attenuation coefficient, is the object distance map in the scene.

[0032] The scattering correction module is configured to perform backward scattering removal and color channel preliminary compensation processing on the image data processed by the preprocessing module based on the physical imaging model and in combination with the parameters output by the physical parameter estimation module. Specifically, the scattering correction module comprises a backward scattering removal submodule and a color channel preliminary compensation submodule. The backward scattering removal submodule calculates the image after removing the backward scattering based on the physical imaging model and in combination with the ambient background light intensity and the initial transmittance provided by the physical parameter estimation module.

[0033] wherein, is the image after removing the backward scattering, is the original image, is the ambient background light intensity, is the initial transmittance, is a small constant to prevent division by zero.

[0034] The color channel preliminary compensation submodule determines the gain of each color channel according to the attenuation degree of the corresponding wavelength of light in water, and performs gain compensation on the image after removing the backscattering by channel, wherein the higher the attenuation degree of the wavelength, the higher the gain of the color channel corresponding to the wavelength.

[0035] The processing of the scattering correction module is equivalent to removing the uniform haze generated by turbid water bodies, effectively improving the image contrast and visible distance. Then, according to the attenuation law of each wavelength of light in water, the gain of each channel of the corrected image is compensated, for example, the gain is increased by channel in the order of blue channel being the lowest, green channel being the second, and red channel being the highest, to preliminarily correct the color distortion phenomenon under water. After scattering correction and color preliminary compensation, the obtained image has significantly improved overall visibility and color.

[0036] The deep learning enhancement module is used to adopt a lightweight deep convolutional neural network to perform local contrast enhancement, brightness adjustment, and color fine correction processing on the image data processed by the scattering correction module, and output an enhanced clear image. The lightweight deep convolutional neural network used by the deep learning enhancement module is specially designed, for example, the U-Net structure, to perform fine enhancement processing on the image output by the previous module (scattering correction module).

[0037] In this embodiment, the lightweight deep convolutional neural network is an encoder-decoder model supporting multi-scale feature extraction, and the specific tasks include local contrast and brightness enhancement, enhancing the contrast of local regions and the overall brightness, and further eliminating residual dark areas.

[0038] Specifically, the encoder adopts a depth separable convolution or group convolution structure to extract multi-scale spatial features of the image; the decoder adopts a bilinear upsampling, fast convolution kernel and lightweight residual block structure to enhance local contrast, adjust brightness and fine correct color; the encoder and the decoder are connected through a jump connection to transfer multi-scale spatial features, and a conditional normalization module is arranged at the network feature fusion layer of the encoder and the decoder. The conditional normalization module introduces the initial transmittance and environmental background light intensity output by the physical parameter estimation module as a conditional constraint to dynamically adjust the normalization parameters and activation function coefficients of the lightweight deep convolutional neural network.

[0039] ​In the embodiment, the embedded image processing unit further comprises a frame rate monitoring module and a display unit module. The frame rate monitoring module is used to monitor the frame rate of the image enhancement processing in real time, and when the frame rate is lower than a preset threshold, the inference rate of the lightweight deep convolutional neural network is automatically adjusted to maintain the processing delay within a preset range. The display unit module is connected with the deep learning enhancement module, used to receive and display the enhanced clear image output by the deep learning enhancement module in real time, and the display unit module supports different resolution display devices according to actual application scenarios.

[0040] Referring to the accompanying drawings Figure 4 and Figure 5 , which are a lightweight deep convolutional neural network structure diagram and a parameter diagram of the embodiment. In Figure 4 , the numbers in the brackets of each level of the network structure are the channel numbers of the level. Figure 5 , which are a lightweight deep convolutional neural network structure diagram and a parameter diagram of the embodiment. In Figure 4 , the numbers in the brackets of each level of the network structure are the channel numbers of the level. Figure 5 , which are a lightweight deep convolutional neural network structure diagram and a parameter diagram of the embodiment. In Figure 4 , the numbers in the brackets of each level of the network structure are the channel numbers of the level.

[0041] In the embodiment, the deep learning part of the underwater image enhancement adopts a lightweight convolutional neural network structure specially designed for the imaging characteristics of turbid water bodies, to achieve the goal of balancing efficiency, interpretability and engineering real-time deployment capability. The network has made targeted innovations in architecture, feature fusion method and parameter control, mainly including the following aspects: Firstly, the network as a whole adopts an encoder-decoder (Encoder-Decoder) framework, which has the ability to extract multi-scale features and reconstruct details from the input image. The encoder part is composed of several layers of depthwise separable convolution (Depthwise Separable Convolution) or group convolution, which significantly reduces the parameter quantity and calculation quantity, while effectively capturing the spatial distribution features and local texture information of the image. After each level of encoder downsampling, the features are transmitted to the symmetric decoder stage through the skip connection (skip connection), which not only helps to fuse high-level semantic information and low-level details, but also effectively prevents information loss caused by multiple convolutions and pooling.

[0042] Secondly, in view of the spatial variability and physical prior of underwater environment imaging, the network specially introduces a conditional normalization (Conditional Normalization) module and a physical parameter fusion channel. In the feature fusion layer of the network, the estimated initial transmittance , background light The physical parameters are encoded as additional input channels and interact with the main feature stream in a conditionally normalized manner at multiple levels of the network. Specifically, the network dynamically adjusts the normalization and activation coefficients of each layer using these physical parameters, thereby guiding the network to adopt adaptive enhancement strategies in different regions. For example, low-transmittance regions (far distance, heavy scattering area) automatically enhance contrast and details, while close-range regions focus more on color restoration and denoising. This mechanism of deeply integrating physical priors into the neural network makes the model highly adaptable to the environment and interpretable, effectively overcoming the drawbacks of traditional deep learning networks, such as "black box" and over-reliance on training distribution.

[0043] Third, to ensure the real-time performance and edge deployment capability of the algorithm, the network fully considers the balance between model size and inference speed during design. On the one hand, techniques such as channel pruning and low-rank decomposition are used to compress the model parameter size, ensuring that the weight size of the entire network is controlled within the level of millions, adapting to the memory and computing power of embedded devices such as Jetson. On the other hand, structures such as bilinear upsampling, fast convolution kernel, and lightweight residual block are used in the decoder part to improve reconstruction efficiency and avoid artifacts and waste of computing power caused by deconvolution in traditional UNet structures.

[0044] In addition, a joint optimization strategy of physics-consistency loss and multi-scale perceptual loss is used during network training. That is, not only is the pixel-level difference (such as L1 / L2 loss) with the enhanced ground truth image used for training, but the network output image is also required to maintain high consistency with the original input image in the physical space after degradation through the physical imaging model. This loss term constrains the network to enhance details without introducing new physical distortions, making the learned enhancement transformation have stronger physical interpretability and generalization ability.

[0045] Finally, the network design fully considers seamless integration with the physical modeling process of the system. The parameters directly output by the physical parameter estimation and image preprocessing stage are embedded in the network input channel and conditionally normalized flow in the form of feature vectors. The output of the enhancement end can be directly displayed or used for subsequent intelligent analysis modules (such as target recognition, three-dimensional reconstruction, etc.), ensuring that the overall architecture of the system is highly coupled and efficient collaboration.

[0046] In the above illustrative embodiment, the lightweight convolutional neural network of the image processing system suitable for turbid water areas greatly reduces the consumption of hardware resources while ensuring that the enhancement effect is significantly improved, meeting all the needs of underwater robots, portable devices and other practical engineering scenarios for efficient, real-time and interpretable image enhancement algorithms. Actual tests show that the single-frame processing time can be controlled within tens of milliseconds, the frame rate is much higher than that of mainstream similar deep learning enhancement networks, and it has good adaptability and stability to different turbid water environments.

[0047] Embodiment 2: Referring to the accompanying drawings Figures 1 to 5 An illustrative embodiment of the image processing method suitable for turbid water areas proposed by the present application is given, which uses the image processing system suitable for turbid water areas of embodiment 1. The image processing method comprises the following steps: S1, real-time acquisition of original image data of turbid water area by underwater camera assembly, and transmission of original image data to embedded image processing unit; S2, noise suppression and image calibration processing of original image data by using pre-processing module to obtain pre-processed image; S3, estimation of environmental background light intensity, water body attenuation coefficient and object distance map in scene of current water area by physical parameter estimation module, and calculation of initial transmittance; S4, based on the physical imaging model, combining the environmental background light intensity and the initial transmittance obtained in step S3, using the scattering correction module to remove the backscattering and preliminarily compensate the color channel of the pre-processed image data, to obtain the scattering corrected image; S5, using lightweight deep convolutional neural network, through deep learning enhancement module to process the scattering corrected image for local contrast enhancement, brightness adjustment and color fine correction, to output the enhanced clear image.

[0048] In some embodiments, in step S3, the specific process of estimating the environmental background light intensity is: analyzing the histogram and color information of the original image data, counting the brightness values of each region in the image, selecting the region with the highest brightness value, and taking the average light intensity of the region as the environmental background light intensity; the specific process of calculating the initial transmittance is: based on the obtained water body attenuation coefficient of the current water area and the object distance in the scene, the initial transmittance is calculated by using the formula wherein, is the initial transmittance, is a natural constant, is the average value of the water body attenuation coefficient, is the object distance map in the scene.

[0049] In some embodiments, in step S4, the specific process of backscattering removal is: based on the physical imaging model, combining the ambient background light intensity and the initial transmittance, the backscattering-removed image is calculated using the formula: wherein, is the backscattering-removed image, is the original image, is the ambient background light intensity, is the initial transmittance, is a small constant to prevent division by zero; the specific process of preliminary compensation of color channels is: according to the attenuation law of light at each wavelength in water, the gain of each color channel is determined according to the attenuation degree of the corresponding wavelength light, and the backscattering-removed image is compensated by gain channel by channel, wherein the gain of the color channel corresponding to the wavelength with higher attenuation degree is set to be higher.

[0050] In some embodiments, in step S5, the training process of the lightweight deep convolutional neural network adopts a joint optimization strategy of physical consistency loss and multi-scale perception loss to train until convergence; the physical consistency loss is used to constrain the output image of the lightweight deep convolutional neural network to maintain consistency with the original input image in the physical space after degradation through the physical imaging model, and the multi-scale perception loss is used to optimize the pixel-level difference between the output image of the lightweight deep convolutional neural network and the target clear image.

[0051] Referring to FIG. 8, Figure 3 is the algorithm flowchart of the present embodiment, first, the system collects the original underwater turbidity image in real time through the underwater camera assembly. To ensure the robustness and accuracy of subsequent processing, the input image will be preliminarily preprocessed by the preprocessing module before entering the main process, such as noise filtering, distortion correction, and resolution unification. After preprocessing, the image is sent to the physical parameter estimation module for the physical parameter estimation process. In this stage, the background light estimation submodule automatically analyzes the overall brightness distribution and color information of the input image, estimates the ambient background light intensity of the current water area by counting the area with the highest brightness in the image , so as to provide basic parameters for subsequent physical modeling.

[0052] Subsequently, the water body attenuation coefficient estimation submodule evaluates the turbidity of the water body. This step usually automatically infers the light attenuation coefficient in water according to prior experience, scene calibration data, or according to the change of color distribution of different channels. Further, in order to more accurately simulate the actual imaging physical process, the system integrates a depth distance estimation submodule for estimating the depth or distance. According to the actual device conditions, the depth distance estimation submodule can automatically call the built-in monocular depth estimation algorithm, binocular disparity measurement, or fuse external ranging information to estimate the distance of each pixel point in the scene to the camera An estimation is made. In combination with the above attenuation coefficient , the initial transmittance calculation submodule automatically converts the distance map into initial transmittance , fully characterizing the attenuation characteristics of light at different depths. If the device is not equipped with an effective depth sensing means, the system automatically switches to the dark channel prior or color distribution based algorithm to approximately calculate the transmittance distribution, ensuring process adaptability and multi-scene applicability.

[0053] After completing the preliminary estimation of physical parameters, the system immediately enters the physical scattering correction phase. At this time, the backscattering removal submodule removes the backscattering component of each pixel point of the original image one by one according to the ambient background light intensity and initial transmittance estimated above. Specifically, the backscattering removal submodule will use the physical model to correct the input image in reverse, removing the uniform haze caused by suspended particles and other factors, thereby significantly improving the contrast and visual clarity of the image. Subsequently, the color channel preliminary compensation submodule automatically compensates for the gain of each color channel of the corrected image according to the attenuation law of light of different wavelengths underwater. Especially for red and green channels that are easily absorbed, the color channel preliminary compensation submodule will automatically increase their intensity according to the empirical ratio to ensure that the overall color of the enhanced image restores the natural visual perception of the real underwater environment as much as possible.

[0054] After the above physical correction and preliminary color compensation, the image will be sent to the deep learning enhancement module for deep learning enhancement process. In this stage, the system takes the physically corrected image and parameters such as initial transmittance and ambient background light as multi-channel input, and feeds them into a specially designed lightweight convolutional neural network. The network first extracts multi-scale spatial features through the encoder, and through mechanisms such as conditional normalization, dynamically embeds physical prior parameters into feature streams, enhancing the network's adaptability to different depths and different environmental regions. The decoder part of the network further restores detail information, improves local contrast, and automatically corrects remaining color distortion. The output end of the network organically fuses the input and the reconstruction result through the residual connection mechanism, so that the enhanced image significantly improves in overall brightness, detail sharpness, color naturalness, and other aspects. For scenes with high real-time requirements, the system also monitors the frame rate of model running through the frame rate monitoring module, automatically adjusts the inference rate to ensure that the processing delay is always maintained within an acceptable range.

[0055] Finally, the system pushes the enhanced high-quality images to the display terminal, i.e., the display unit module, with minimal delay for real-time viewing or analysis by the underwater robot operator, scientific researchers or automated recognition systems. The entire process from image acquisition, parameter estimation, physical modeling correction, deep learning enhancement to result output is fully automated without human intervention, and the algorithm structure is highly modular and can be optimized in parallel, which can also run smoothly on limited resource platforms such as embedded or on-board GPUs, fully meeting the real-time and robustness requirements in actual underwater operating environments.

[0056] In the above exemplary embodiments, the image processing method suitable for turbid water areas has been experimentally verified to have excellent enhancement effect and practical value: first, the clarity is significantly improved and the underwater visibility range is expanded. In a highly turbid environment that is almost invisible to the naked eye, the enhanced image clearly presents the scene details and targets, solving the problem of "unclear" for ordinary cameras. The object contour and texture can be distinguished, and compared with the traditional underwater dark channel prior (UDCP) method, the output image has higher contrast and fewer fogging artifacts in low-visibility scenes, and can also effectively identify hidden objects in the distance; second, the color is restored realistically and the image is natural and true. With the constraint and learning correction of the physical model, the color of the enhanced image is accurate, avoiding strange color tones. Compared with traditional methods such as CLAHE and direct GAN enhancement, the color is closer to the human eye's perception in clear water environment, such as white objects still showing true white, which is of great significance for underwater scientific research and engineering applications; third, multiple problems can be solved integrally and the overall image quality can be improved. A set of algorithms simultaneously processes multiple underwater image degradation problems such as defogging, brightening, denoising and sharpening, so that the output image reaches a relatively optimal level in clarity, brightness and details. Through module cooperation, the side effects such as noise enhancement and detail loss caused by single optimization in existing solutions are avoided; fourth, it has excellent real-time performance and deployment convenience. After optimization, it can run in real time on an embedded platform, with millisecond-level delay to meet the task requirements of autonomous underwater vehicle (AUV) navigation, ROV operation, etc. On an embedded GPU platform, it can achieve more than 30 frames per second of enhancement processing, and mainly relies on software implementation without the need for special optical devices or multi-camera configuration, with strong portability and low deployment cost; fifth, it has strong adaptability and broad application prospects. By integrating physical models and adaptive learning strategies, it is suitable for different types of water areas such as seawater and lake water, as well as different turbidity and depth conditions. Once deployed, it can cope with multiple environments and provide clear and reliable image information for marine scientific research, industrial detection, underwater safety rescue and other scenarios, significantly improving the target discovery and recognition success rate.

[0057] Finally, it should be noted that: the embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other.

[0058] The above examples are only used to illustrate the technical solutions of the present application but not to limit the present application; although the present application has been described in detail with reference to the preferred embodiments, it is understood by the person of ordinary skill in the art that the specific embodiments of the present application can be modified or some technical features can be replaced by equivalent ones without departing from the spirit of the technical solutions of the present application, and all of them should be covered in the technical solution range of the present application claimed.

Claims

1. An image processing system suitable for use in turbid water, characterized in that, The underwater camera assembly and the embedded image processing unit are included; The underwater camera assembly is used for collecting raw image data of turbid water in real time and transmitting the collected raw image data to the embedded image processing unit; The embedded image processing unit includes a pre-processing module, a physical parameter estimation module, a scattering correction module and a deep learning enhancement module connected in sequence; The pre-processing module is used for noise suppression and image calibration processing of the raw image data; The physical parameter estimation module is used for estimating the environmental background light intensity, water body attenuation coefficient and object distance map in the scene of the current water area, and calculating the initial transmittance; The scattering correction module is used for removing backward scattering and performing preliminary compensation of color channels on the image data processed by the pre-processing module based on a physical imaging model and in combination with the parameters output by the physical parameter estimation module; The deep learning enhancement module is used for performing local contrast enhancement, brightness adjustment and color fine correction processing on the image data processed by the scattering correction module by using a lightweight deep convolutional neural network, and outputting an enhanced clear image.

2. The image processing system suitable for turbid water according to claim 1, characterized in that, The physical parameter estimation module includes a background light estimation submodule, a water body attenuation coefficient estimation submodule, a depth distance estimation submodule and an initial transmittance calculation submodule; The background light estimation submodule estimates and determines the environmental background light intensity by analyzing the histogram and color information of the image data and selecting the area with the lightest color or the highest brightness in the image; The water body attenuation coefficient estimation submodule determines the water body attenuation coefficient of each color channel by evaluating the turbidity of the water body; The depth distance estimation submodule estimates the object distance map in the scene by using a pre-trained monocular depth neural network, binocular camera ranging and fusion of external ranging information; The initial transmittance calculation submodule calculates the initial transmittance based on the water body attenuation coefficient and the object distance map: wherein, is the initial transmittance, is a natural constant, is the average value of the water attenuation coefficient, is the object distance map in the scene.

3. The image processing system suitable for turbid water according to claim 1, wherein, The scattering correction module includes a backward scattering removal submodule and a color channel preliminary compensation submodule; The backward scattering removal submodule calculates the image after removing backward scattering based on a physical imaging model and in combination with the environmental background light intensity and the initial transmittance: wherein, is the image after removal of backscattering, is the original image, is the ambient background light intensity, is the initial transmittance, is a small constant to prevent division by zero; The color channel preliminary compensation submodule determines the gain level of each color channel according to the attenuation degree of the corresponding wavelength light according to the attenuation law of each wavelength light in water, and performs gain compensation on the image after removing backward scattering, wherein the gain of the color channel corresponding to the wavelength with higher attenuation degree is set to be higher.

4. The image processing system suitable for turbid water according to claim 1, wherein, The lightweight deep convolutional neural network is an encoder-decoder model, the encoder adopts a deep separable convolution or group convolution structure to extract multi-scale spatial features of the image; the decoder adopts a bilinear upsampling, a fast convolution kernel and a lightweight residual block structure to improve local contrast, adjust brightness and finely correct color; the multi-scale spatial features are transmitted between the encoder and the decoder through a skip connection, and a conditional normalization module is arranged at a network feature fusion layer of the encoder and the decoder, the conditional normalization module introduces the initial transmittance and the ambient background light intensity output by the physical parameter estimation module as a conditional constraint to dynamically adjust normalization parameters and activation function coefficients of the lightweight deep convolutional neural network.

5. The image processing system suitable for turbid water according to claim 1, wherein, The embedded image processing unit further comprises a frame rate monitoring module, which is used to monitor the frame rate of image enhancement processing in real time, and automatically adjust the inference rate of the lightweight deep convolutional neural network when the frame rate is lower than a preset threshold, so that the processing delay is maintained within a preset range.

6. The image processing system suitable for turbid water according to claim 1, wherein, The embedded image processing unit further comprises a display unit module connected with the deep learning enhancement module, which is used to receive and display the enhanced clear image output by the deep learning enhancement module in real time, and the display unit module supports different resolution display devices according to actual application scenarios.

7. An image processing method suitable for turbid water, characterized by, The image processing system suitable for turbid water areas according to any one of claims 1-6, the image processing method comprises the following steps: S1, real-time acquisition of original image data of turbid water area by underwater camera assembly, and transmission of original image data to embedded image processing unit; S2, noise suppression and image calibration processing of original image data by pre-processing module to obtain pre-processed image; S3, estimation of ambient background light intensity, water body attenuation coefficient and object distance map in scene of current water area by physical parameter estimation module, and calculation of initial transmittance; S4, based on the physical imaging model, combining the ambient background light intensity and the initial transmittance obtained in step S3, using the scattering correction module to remove the backscattering and preliminarily compensate the color channel of the pre-processed image data, to obtain the scattering corrected image; S5, using a lightweight deep convolutional neural network, the scattering corrected image is processed by a deep learning enhancement module to improve local contrast, adjust brightness and finely correct color, and output an enhanced clear image.

8. The image processing method suitable for turbid water according to claim 7, wherein, In step S3, the specific process of estimating the ambient background light intensity is: analyzing the histogram and color information of the original image data, counting the brightness values of each region in the image, selecting the region with the highest brightness value, and taking the average light intensity of the region as the ambient background light intensity; the specific process of calculating the initial transmittance is: based on the obtained water body attenuation coefficient of the current water area and the object distance in the scene, the initial transmittance is calculated by using the formula wherein, is the initial transmittance, is a natural constant, is the average value of the water body attenuation coefficient, is the object distance map in the scene.

9. The image processing method suitable for turbid water according to claim 7, wherein, In step S4, the specific process of backscattering removal is: based on the physical imaging model, combining the environmental background light intensity and the initial transmittance, the backscattering-removed image is calculated by using the formula: wherein, is the backscattering-removed image, is the original image, is the environmental background light intensity, is the initial transmittance, is a small constant for preventing division by zero; the specific process of the preliminary compensation of the color channels is: according to the attenuation law of light rays of each wavelength in water, the gain of each color channel is determined according to the attenuation degree of the corresponding wavelength light, and the backscattering-removed image is compensated by gain channel by channel, wherein the gain of the color channel corresponding to the wavelength with the higher attenuation degree is set to be higher.

10. The image processing method suitable for turbid water according to claim 7, wherein, In step S5, the training process of the lightweight deep convolutional neural network adopts a joint optimization strategy of physical consistency loss and multi-scale perception loss to train until convergence; the physical consistency loss is used to constrain the output image of the lightweight deep convolutional neural network to maintain consistency with the original input image in the physical space after degradation through the physical imaging model, and the multi-scale perception loss is used to optimize the pixel-level difference between the output image of the lightweight deep convolutional neural network and the target clear image.

Images