An underwater image recovery method and system for a low-computing-power embedded platform

Abstract

The application belongs to the technical field of underwater image processing, and relates to an underwater image recovery method and system for a low-computing-power embedded platform. The method comprises the following steps: collecting an original underwater degraded image; preprocessing the image to obtain a standardized image and extracting a basic physical feature set; constructing a scattering indication feature set; estimating a scattering component through a light neural network; compensating the standardized image according to the scattering component; recovering visible information of a target weakened by scattering based on a physical constraint to obtain a recovered image; obtaining a recovery confidence weight based on the scattering indication feature set and normalized scattering intensity, and correcting the recovered image; and post-processing the corrected recovered image to output a final recovered image. Under the condition of a low-computing-power embedded platform, the application preferentially concentrates limited computing power on solving a scattering component rather than enhancing an entire image, and then realizes effective recovery of underwater image information through scattering compensation, physical constraint recovery and recovery confidence weight correction.

Description

Technical Field

[0001] This invention belongs to the field of underwater image processing technology, specifically relating to an underwater image restoration method and system for low-computing-power embedded platforms. Background Technology

[0002] Underwater images differ from those in ordinary air. The biggest problem is not simply "darkening" or "color distortion," but rather that water itself significantly absorbs and scatters light. Absorption leads to inconsistent color attenuation, while scattering causes images to appear grayish, with reduced contrast and obscured details. In severe cases, the target itself may still be in the image, but the effective information has been obscured by the scattered background, making it appear as if it is hidden. This problem is particularly pronounced in highly turbid waters, near-field lighting, engineering waters, ponds, ports, and inland rivers.

[0003] Currently, underwater image processing methods can be broadly categorized as follows: The first category is traditional image enhancement methods, such as white balance, contrast stretching, histogram equalization, sharpening, Retinex, and color correction. The advantages of these methods are their simplicity, relatively low computational cost, and ease of deployment on devices. However, the main problem is that these methods mostly only visually enhance the final image without actually calculating why the underwater image degrades or separating the interference caused by water scattering. Therefore, when scattering is severe, these methods can usually only make the image "brighter" or "clearer," but it's difficult to truly recover the effective information obscured by scattering, and sometimes they even amplify noise and artifacts.

[0004] The second category is underwater image restoration methods based on physical models. These methods typically first establish an underwater imaging model, then estimate parameters such as background light, transmittance, and attenuation coefficients, and finally deduce the target image. The advantage of this type of method is that it is more interpretable than simple image enhancement and theoretically closer to the actual imaging process; however, it has the following problems: 1) It requires relying on relatively strong prior assumptions, such as assuming that the water body is relatively uniform, assuming that the scene depth changes relatively smoothly, and assuming that certain areas can represent background light, etc. 2) In real, complex waters, scattering is often spatially variable and cannot be described by a fixed parameter. In particular, when factors such as forward scattering, back scattering, local suspended particles, and changes in the position of supplementary lighting are superimposed, it is difficult to rely solely on traditional formulas to calculate accurately and stably. 3) It is better at handling overall fogging and overall color distortion, but its recovery ability is often not stable enough when "local target information is scattered and blocked".

[0005] The third category comprises underwater image enhancement or restoration methods based on deep learning, which have become increasingly popular in recent years. These methods typically use convolutional neural networks or other network structures to directly map degraded images into enhanced images, or output results such as transmittance, background light, and sharpness. The advantages of this type of method are its strong expressive power and relatively good performance on some public datasets; however, it suffers from the following problems: 1) Essentially, it is still "end-to-end enhancement", which is more focused on visual effect optimization and may not actually model scattering separately; 2) Many networks are not small in scale and have requirements for computing power, video memory, storage and power consumption, which are not suitable for long-term real-time operation of low computing power embedded devices; 3) If the water environment and training data are significantly different, the model is prone to undergeneralization. The image may look "better" but it may not actually recover the occluded information.

[0006] Other image processing methods are starting to focus on lightweight deployment, such as building small networks, real-time enhancement, and edge inference. However, they still use limited computing power for the enhancement, correction, or style restoration of the entire image. In other words, the network is "doing a little bit of everything" instead of concentrating limited computing power on the most critical and difficult parts. From the perspective of underwater imaging mechanisms, the most difficult part and the part that affects the visibility of the target is actually the scattering component, especially the part of the scattering effect that changes dynamically with the water state, target distance, local illumination, and suspended particle concentration. If this part is not calculated accurately, the subsequent transmission restoration, contrast enhancement, and detail reconstruction will all be affected.

[0007] In summary, the common problems of existing underwater image processing methods can be summarized as follows: some methods have low computational cost, but only perform surface enhancement and do not truly solve the problem of information occlusion caused by scattering; some methods have physical models, but their descriptions of complex, non-uniform, and dynamically changing scattering are not accurate enough; some methods rely on deep networks, which seem to have good results, but the models are too large, require high computing power, and tend to be biased towards overall visual enhancement, lacking clear physical interpretation and unsuitable for long-term real-time operation on low-computing-power embedded platforms; existing lightweight image processing methods do not concentrate computing power on the most difficult-to-solve and most impactful scattering components that affect target visibility, but instead distribute computing power across the entire image enhancement, resulting in insufficient recovery efficiency and specificity under limited computing power. Summary of the Invention

[0008] To address the shortcomings of related technologies, this invention provides an underwater image restoration method and system for low-computing-power embedded platforms, aiming to achieve effective restoration of underwater image information under conditions of limited computing power.

[0009] This invention provides an underwater image restoration method for low-computing-power embedded platforms, comprising the following steps: S1. Acquire raw underwater degradation images , Indicates the pixel position in the image; S2, to Preprocessing is performed to obtain a standardized image. ;from Extract the fundamental physical features and denote their set as follows: ; S3, from Extract scattering-related features and construct a scattering indicator feature set. ; S4. Estimating scattering components using a lightweight neural network. , expressed as equation (8); where, This represents a lightweight neural network. Represents network parameters, Indicates network input characteristics, ;right Normalization is performed to obtain normalized scattering intensity. ; (8); S5. Based on the estimated scattering components Standardized images Perform scattering compensation to obtain a scattering-compensated image. ; S6. Recover the target's visibility information weakened by scattering based on physical constraints to obtain the recovered image. , expressed as equation (14); where, This represents the estimated background light component. This represents the estimated transmittance. Indicates a stable term; (14); S7, Based on scattering indicator feature set and normalized scattering intensity Obtain the recovery confidence weight ; and then according to equation (20) Make corrections to obtain the corrected restored image. ; (20); S8, to Post-processing is performed to output the final restored image. .

[0010] In some embodiments, in step S3, a scattering indicator feature set is constructed according to equation (1). ;in, This represents the scattering indicator feature construction function. , , , , They represent Local average brightness, local contrast, color channel attenuation differences, edge response, and local blur degree; (1).

[0011] In some embodiments, in step S4, the lightweight neural network Employs a lightweight encoder-decoder architecture. It consists of an input convolutional layer, several depthwise separable convolutional layers, a lightweight residual block, and an output convolutional layer. The input convolutional layer performs channel mapping; the depthwise separable convolutional layers extract local scattering features and reduce computational cost; the lightweight residual block enhances feature representation and preserves details; and the output convolutional layer outputs the estimated scattering components. .

[0012] In some embodiments, in step S5, a scattering compensation image is obtained according to equation (11). ,in, The scattering compensation weight is adaptively generated based on local scattering characteristics. (11).

[0013] In some embodiments, in step S5, scattering compensation weights are generated according to equation (12). ;in, Indicates according to A function for adaptively calculating scattering compensation weights. This represents the truncation function. express Normalized local contrast , , For adjustment coefficients, and These represent the lower and upper limits of the scattering compensation weights, respectively. (12).

[0014] In some embodiments, in step S6, Based on scattering compensation image In pixel position local window centered The low-frequency brightness or color statistics within the area are estimated using methods such as local mean, weighted average, or local bright candidate region mean. It is the normalized scattering intensity and The normalized local contrast is combined with the truncation function for estimation.

[0015] In some embodiments, in step S7, the recovery confidence weights are constructed according to equation (21). ;in, Indicates according to and A function for calculating the recovery confidence weights. This represents the truncation function. express Normalized local contrast express Normalized edge response, , , Represented as weighting coefficients, ; (twenty one).

[0016] The present invention also provides an underwater image restoration system for low-computing-power embedded platforms, used to perform the aforementioned underwater image restoration method for low-computing-power embedded platforms, the underwater image restoration system for low-computing-power embedded platforms comprising: Image acquisition module, used to acquire raw underwater degraded images; The preprocessing module performs preprocessing operations on the original underwater degraded image, including resizing, color space transformation, and normalization, to obtain a standardized image and extract a set of basic physical features. The scattering indicator feature construction module is used to extract scattering-related features from the normalized image and construct a set of scattering indicator features; A lightweight neural network scattering estimation module for estimating scattering components in standardized images under low computational power conditions; The physical constraint recovery module is used to compensate the normalized image based on the estimated scattering components to obtain a scattering compensation image, and then recover the target visibility information weakened by scattering based on physical constraints to obtain a recovery image. Finally, the recovery image is corrected using recovery confidence weights. The post-processing output module performs post-processing operations on the corrected restored image, including detail smoothing, color consistency correction, brightness range constraint, and output format conversion, to output the final restored image.

[0017] Based on the above technical solutions, the underwater image restoration method and system for low-computing-power embedded platforms in this invention concentrates the limited computing power of lightweight neural networks on the adaptive solution of underwater scattering components rather than end-to-end enhancement of the entire image. That is, under low computing power conditions, it prioritizes solving the key problem of scattering, and then improves the target visibility and restoration stability in the scatter-weakened area through scattering compensation, physical constraint restoration, and restoration confidence weight correction. Compared with the existing underwater image enhancement, ordinary dehazing, and lightweight end-to-end network solutions, this invention has better application value in terms of underwater image restoration effect, stability, interpretability, and engineering feasibility. Attached Figure Description

[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, illustrate exemplary embodiments of the invention and, together with their description, serve to explain the invention and do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart of the underwater image restoration method for low-computing-power embedded platforms according to the present invention; Figure 2 This is a rendering of the underwater image restored using existing technology. Figure 3 This is a diagram showing the effect of underwater image restoration obtained using the present invention. Detailed Implementation

[0019] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0020] In the description of this invention, it should be understood that the terms "center", "lateral", "longitudinal", "upper", "lower", "top", "bottom", "inner", "outer", "left", "right", "front", "rear", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0021] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0022] refer to Figure 1 As shown, this invention provides an underwater image restoration method for low-computing-power embedded platforms. It is applicable to low-computing-power embedded platforms such as underwater cameras, underwater robots, underwater observation equipment, and portable underwater imaging terminals, and is particularly suitable for real-time image restoration in turbid waters, non-uniform scattering environments, and close-range supplemental lighting environments. The underwater image restoration method for low-computing-power embedded platforms of this invention includes the following steps S1 to S8.

[0023] Step S1: Acquire raw underwater degradation images: Acquire raw underwater degradation images using equipment such as underwater cameras, underwater robots, underwater observation equipment, or portable underwater imaging terminals, and record them as follows: ,in, This indicates the pixel position in the image.

[0024] Step S2: Preprocess the original underwater degraded image: Preprocess the original underwater degraded image. Preprocessing operations such as resizing, color space transformation, and normalization are performed to obtain a standardized image. The specific methods of preprocessing are well known to those skilled in the art and will not be elaborated here; from standardized images Extract the fundamental physical features and denote their set as follows: It is understandable that the set of fundamental physical characteristics This includes brightness, local contrast, color attenuation, texture intensity, and edge response.

[0025] Step S3: Constructing a scattering indicator feature set: Since scattering in underwater images typically manifests as local brightness enhancement, local contrast reduction, increased differences in color channel attenuation, blurred details, edge diffusion, or weakened edge response, this step constructs a scattering indicator feature set. It is expressed as Equation (1), which is derived from the standardized image. Features related to scattering are extracted to characterize the intensity and spatial distribution of underwater scattering; among them, This represents the scattering indicator feature construction function. Represents a standardized image The local average brightness is used to reflect the local brightness increase caused by the scattered background. Represents a standardized image The local contrast is used to reflect the reduction in local grayscale differences caused by scattering; Represents a standardized image The color channel attenuation difference is used to reflect the color shift caused by the different absorption and scattering degrees of different color channels in water; Represents a standardized image The edge response is used to reflect edge weakening or edge diffusion caused by scattering; Represents a standardized image The degree of local blurring is used to reflect the degree of degradation of image details; (1).

[0026] To further explain, , , , , Calculations are performed according to equations (2) to (6) respectively; where, , , Representing the standardized images At pixel position The red channel value, green channel value, and blue channel value at that location. and Represents a standardized image luminance component Gradient response in the horizontal and vertical directions, Represents a standardized image In terms of pixel position A local window centered on the center. This indicates the number of pixels within the local window. This represents the summation index variable. Represents the luminance component The Laplace response; (2); (3); (4); (5); (6).

[0027] Those skilled in the art will understand that if standardized images For single-channel images such as grayscale images, the luminance component This refers to the brightness value of a single-channel image; if the image is standardized... This is a three-channel RGB image, with the luminance component... Calculate according to equation (7); (7).

[0028] Step S4: Estimate the scattering components in the image using a lightweight neural network: In this step, the lightweight neural network is not used to directly output the final restored image, but is specifically used to estimate the scattering components, which are the most difficult to explicitly model during underwater image degradation; scattering components The estimation is performed according to equation (8); where, This represents a lightweight neural network. Represents network parameters, The input features of the lightweight neural network are represented by equation (9); (8); (9).

[0029] It is composed of the aforementioned standardized image itself, the basic physical feature set, and the artificially constructed scattering indicator feature set. That is, this embodiment does not simply input the original image directly into the lightweight neural network, but first constructs a scattering indicator feature set related to underwater scattering (which can reflect scattering-related phenomena such as local brightness increase, local contrast decrease, color channel attenuation difference, edge weakening, texture weakening, and local blurring), and then uses it together with the image as network input. By explicitly adding scattering indicator features at the network input, the learning pressure of the lightweight neural network on the low computing power embedded platform is reduced, the difficulty of the lightweight neural network to learn the scattering discrimination criteria on its own under low computing power conditions is reduced, and the pertinence and stability of scattering component estimation under low computing power conditions are improved.

[0030] To further explain, lightweight neural networks A lightweight encoder-decoder architecture can be used. It consists of an input convolutional layer, several depthwise separable convolutional layers, a lightweight residual block, and an output convolutional layer. The input convolutional layer performs channel mapping; the depthwise separable convolutional layers extract local scattering features and reduce computational cost; the lightweight residual block enhances feature representation and preserves details; and the output convolutional layer outputs the estimated scattering components. . It can be the scattering component in a single-channel image or the three-channel scattering component corresponding to an RGB image; typically, the output convolutional layer uses 1x1 convolution for channel mapping and constrains the output to the range of 0 to 1 using the Sigmoid function.

[0031] Furthermore, when When the scattering component is in a single-channel image, it can be directly normalized and used as the normalized scattering intensity. .when When dealing with the three-channel scattering components corresponding to an RGB image, the normalized scattering intensity can be obtained using the channel mean, channel maximum, or luminance-weighted method. As shown in equation (10), where, , , These represent the three-channel scattering components; normalized scattering intensity. Used for subsequent calculations of scattering compensation weights, transmittance, and recovery confidence weights; (10).

[0032] Step S4 clarifies that under low computing power conditions, the limited computing power of the neural network should be prioritized for the key task of solving the scattering components. This makes the network task boundary clearer, more targeted, and more suitable for real-time operation on low computing power embedded platforms. Although some lightweight underwater image processing methods now use small neural networks, they usually allow the network to directly enhance, correct, or restore the entire image end-to-end. This would disperse the limited computing power across multiple tasks such as brightness adjustment, color correction, contrast enhancement, and texture reconstruction, making it difficult to prioritize solving the scattering problem, which has the greatest impact on the visibility of underwater targets.

[0033] Step S5: Generate a scattering compensation image based on the scattering estimation results: based on the estimated scattering components... Standardized images Perform scattering compensation to obtain a scattering-compensated image. , expressed as equation (11); where, This represents the scattering compensation weight, which is adaptively generated based on local scattering characteristics; (11).

[0034] To further explain, the scattering compensation weight The calculation is performed according to equation (12); where, Represents the set of scattering indication features A function for adaptively calculating scattering compensation weights. This represents the truncation function. Represents a standardized image Normalized local contrast, by normalizing the image Local contrast The result is obtained after normalization. , , The adjustment coefficients control the base weight, scattering intensity term weight, and local contrast term weight during compensation, respectively. and These represent the lower and upper limits of the scattering compensation weights, respectively; Record , Indicates the variable Limited to to Within the range; Equation (12) indicates that the stronger the scattering intensity and the lower the local contrast, the greater the scattering compensation weight; when the scattering intensity is weak and the local contrast is high, the smaller the scattering compensation weight; thus, the adaptive generation of the scattering compensation weight is achieved, avoiding undercompensation or overcompensation. (12).

[0035] As can be seen from step S5, scattering compensation is not simply subtracting a fixed scattering estimate from the entire image, but rather setting scattering compensation weights so that the compensation intensity can adaptively change according to the local scattering state. This allows for stronger compensation in high-scattering areas and reduced compensation intensity in low-scattering areas, thereby adapting to spatial variations in non-uniform turbid water bodies and improving the recovery stability of different areas.

[0036] Step S6: Recover the target visibility information weakened by scattering based on physical constraints to obtain the recovered image.

[0037] It should be noted that in this embodiment, the degradation process of underwater images is represented in the general form shown in equation (13), wherein, Represents the observed image, This represents the target image to be recovered. Indicates transmittance. Indicates the background light component. Indicates the scattering component; (13).

[0038] In this degradation process, this embodiment involves the scattering component. The scattering components are isolated from the overall image degradation process and treated as an independent solution object. Furthermore, the scattering components are not directly calculated using fixed empirical formulas, but rather adaptively estimated using a lightweight neural network in step S4 to obtain the estimated scattering components. Then, the scattering compensation image is obtained through step S5. Then based on the scattering compensation image Perform physical constraint recovery, i.e., combine the estimated background light components. and estimated transmittance Obtain the restored image , expressed as equation (14); where, This indicates a stable term, used to avoid the denominator being too small; (14); It should be noted that equation (14) is based on the scattering compensation image. Restoration will be carried out on the basis of existing practices; due to The estimated scattering components have been subtracted or weakened according to equation (11). Therefore, in equation (14), no further deduction is required. This is to avoid duplicate compensation.

[0039] To further explain, Based on scattering compensation image In pixel position local window centered The low-frequency brightness or color statistics within the region are estimated using methods such as local mean, weighted average, or local bright candidate region mean; for example, the local mean method shown in equation (15) can be used to calculate... ,in, Represents the scattering compensation image In terms of pixel position A local window centered on the center. This indicates the number of pixels within the local window. Indicates the summation index variable; (15).

[0040] To further explain, It is the normalized scattering intensity and The normalized local contrast is combined with a cutoff function for estimation; the estimated transmittance... The calculation is performed according to equation (16), where, Represents the scattering compensation image Normalized local contrast, through scattering compensation image Local contrast The result is obtained after normalization. , The adjustment coefficients control the weights of the scattering intensity term and the local contrast term, respectively. Indicates the lower limit of transmittance; Record , Indicates the variable Limited to Within the range of 1; Equation (16) indicates that the stronger the scattering intensity and the lower the local contrast, the smaller the estimated transmittance value; when the scattering intensity is weak and the local contrast is high, the estimated transmittance value is large. (16).

[0041] Furthermore, scattering compensation image Local contrast The calculation is performed according to equation (17), where, Represents the scattering compensation image In terms of pixel position A local window centered on the center. This indicates the number of pixels within the local window. This represents the summation index variable. Represents the scattering compensation image The brightness component, Represents the scattering compensation image The local average brightness is calculated according to Equation (18). It can be understood that the calculation principle of the local average brightness represented by Equation (18) and Equation (2) is the same, and the calculation principle of the local contrast represented by Equation (17) and Equation (3) is also the same, except that the image properties they are used for calculation are different. (17); (18); Then, according to equation (19), Normalization is performed to obtain the scattering compensation image. Normalized local contrast ,in, It is a positive stable term; (19).

[0042] As can be seen from step S6, this embodiment does not mix all degradation factors for processing. Instead, it first explicitly estimates and compensates for the scattering component, and then uses a physical constraint model to process the background light, transmittance, and image restoration process, which has clear interpretability. This restores or enhances the visible information of the target weakened by scattering, making it more targeted and interpretable. In contrast, many existing underwater image enhancement or restoration methods involve concepts such as scattering, transmittance, and background light, but in actual implementation, they often mix these degradation factors in the overall enhancement or end-to-end restoration process without explicitly estimating the scattering component as an independent object. Therefore, this embodiment avoids the excessive reliance on fixed priors of pure physical methods and the excessive requirements of pure end-to-end networks on computing power and data scale, thus balancing restoration effect, model interpretability, and low-computing-power deployment capability.

[0043] Step S7: Construct recovery confidence weights and correct image restoration results: Due to the possibility of local reflections, extremely dark areas, missing textures, uneven lighting, or weakened target edges in underwater scenes, even if a lightweight neural network can estimate the scattering components, there may still be risks of miscompensation, over-enhancement, or local distortion in individual areas; therefore, this step is based on the scattering indicator feature set. and normalized scattering intensity Obtain the recovery confidence weight ; and then, according to equation (20), the recovered image obtained in step S6 is processed. Make corrections to obtain the corrected restored image. Equation (20) indicates that when When the image quality is high, image restoration is more commonly used. ;when At lower levels, more standardized images are preserved. This reduces the risk of false recovery, over-enhancement, and local artifacts; (20).

[0044] To further explain, the recovery confidence weights are constructed according to equation (21). ;in, Represents the set of scattering indication features and normalized scattering intensity A function for calculating the recovery confidence weights. This represents the truncation function. Represents a standardized image Normalized local contrast, by normalizing the image Local contrast The result is obtained after normalization. Represents a standardized image The normalized edge response is obtained by normalizing the image. Edge response The result is obtained after normalization. , , These are represented as weighting coefficients, which control the impact of scattering intensity, local contrast, and edge response on the confidence score, respectively. ;Will Record , Indicates the variable The range is limited to 0 to 1; Equation (21) indicates that when the scattering intensity is strong, the local contrast is low, and the edge response is weak, it means that the region is more likely to be scattered and the recovery confidence weight is higher; otherwise, the recovery confidence weight is lower.

[0045] (twenty one).

[0046] Furthermore, standardized images Edge response Calculations are performed according to equation (5), and then according to equation (22)... Normalization is performed to obtain a standardized image. Normalized edge response ,in, It is a positive stable term; it is understandable that in equation (21) It was also obtained based on a similarity normalization process; (twenty two).

[0047] Step S7 further considers how to control the recovery intensity and recovery risk in complex areas, thereby making the entire underwater image recovery process more stable.

[0048] Step S8: The corrected and restored image obtained in step S7. Post-processing operations such as detail smoothing, color consistency correction, brightness range constraint, and output format conversion are performed to output the final restored image. The specific methods for post-processing are well known to those skilled in the art and will not be elaborated here.

[0049] The above illustrative embodiment explicitly uses a lightweight neural network for scattering component estimation rather than whole-image enhancement, and explicitly adds a scattering indicator feature set to the network input to help improve the network's scattering estimation capability, greatly reducing the learning pressure of small networks under low computing power conditions; it adopts a two-stage recovery link of "first solving for scattering, then physical recovery", concentrating the limited computing power of the neural network on solving the scattering components that are most difficult to accurately model in the underwater imaging degradation process, and then recovering or enhancing the target visibility information weakened by scattering through physical constraints, and reducing the risk of false recovery, over-enhancement and local artifacts through a recovery confidence weight mechanism; therefore, this embodiment can achieve reliable and effective recovery of underwater image information under low computing power embedded platform conditions; Reference Figure 2 , Figure 3 As shown in the experiment, compared with the commonly used underwater enhancement, ordinary defogging, or pure end-to-end small network solutions in the existing technology, this embodiment can significantly improve the recovery effect of underwater image information and is more suitable for real-time image recovery in turbid water, non-uniform scattering environment, and close-range supplementary lighting environment.

[0050] This invention also provides an underwater image restoration system for low-computing-power embedded platforms, used to execute the aforementioned underwater image restoration method for low-computing-power embedded platforms. The underwater image restoration system includes an image acquisition module, a preprocessing module, a scattering indication feature construction module, a lightweight neural network scattering estimation module, a physical constraint restoration module, and a post-processing output module.

[0051] The image acquisition module acquires raw, degraded underwater images. The preprocessing module performs preprocessing operations on the raw underwater images, including resizing, color space transformation, and normalization, to obtain a standardized image and extract a set of basic physical features (i.e., a general set of image physical features, including brightness features, local contrast features, color attenuation features, texture intensity features, edge response features, etc.). The scattering indicator feature construction module extracts scattering-related features from the standardized image (i.e., features that help determine the strength and spatial distribution of scattering, such as local average brightness features, local contrast features, color channel attenuation difference features, edge response features, local blurring degree features, etc.) and constructs a set of scattering indicator features for subsequent scattering component estimation. It is understood that the basic physical feature set and the scattering indicator feature set may contain some of the same or related low-level image features, but the role of the scattering indicator feature set is to be specifically constructed for scattering component estimation. The lightweight neural network scattering estimation module is used to estimate the scattering components in the standardized image under low computational power conditions. The physical constraint restoration module compensates the normalized image based on the estimated scattering components to obtain a scattering-compensated image. It then restores the target visibility information weakened by scattering based on physical constraints to obtain a restored image, which is further corrected using restoration confidence weights. The post-processing output module performs post-processing operations on the corrected restored image, including detail smoothing, color consistency correction, brightness range constraints, and output format conversion, to output the final restored image.

[0052] The above illustrative embodiment mainly constructs a two-stage recovery link through the setting of a lightweight neural network scattering estimation module and a physical constraint recovery module. Specifically, the lightweight neural network scattering estimation module is responsible for solving the scattering components that are most difficult to explicitly analyze, while the physical constraint recovery module is responsible for completing the image information recovery using an interpretable model. This forms a two-stage collaborative mechanism of "scattering estimation + physical recovery," enabling effective recovery of underwater image information based on a low-computing-power embedded platform. Compared with the pure physical methods or pure end-to-end methods in the prior art, this embodiment can run without relying on large models, large video memory, or high-power platforms, has clear engineering applicability, and is more suitable for practical engineering deployment.

[0053] Through the description of several embodiments of the underwater image restoration method and system for low-computing-power embedded platforms of the present invention, it can be seen that the present invention has at least one or more of the following advantages: 1) Effective underwater image restoration can be achieved on low-computing-power embedded platforms: Existing methods either focus on image enhancement and find it difficult to truly restore information obscured by scattering; or they rely on large models and high computing power, which are not suitable for real-time embedded operation; while this invention prioritizes the limited computing power of lightweight neural networks on low-computing-power embedded platforms to solve the scattering components that have the greatest impact on the visibility of underwater targets rather than the whole image enhancement, and can still achieve good restoration results under low computing power conditions; 2) More targeted treatment of underwater scattering problems: This invention extracts and prioritizes the scattering component from the overall degradation process, and then compensates and restores it accordingly. This is different from the existing technology that mixes multiple degradation factors, and is therefore more conducive to solving the information masking problem in complex and turbid environments. 3) It can balance recovery capability, interpretability and engineering feasibility: The present invention adopts a collaborative technical route of "scattering estimation + physical recovery". It uses lightweight neural networks to handle complex scattering problems and physical constraints to recover subsequent image information. This avoids the excessive reliance of pure physical methods on fixed priors and avoids the excessive requirements of pure end-to-end networks on computing power and data scale, and is more suitable for actual engineering deployment. 4) Improved scattering estimation performance under low computing power conditions: This invention constructs a set of scattering indicator features to guide a lightweight neural network to focus on scattering-related information, thereby reducing the difficulty for a small network to learn complex scattering laws on its own, enhancing the scattering estimation capability of the lightweight neural network under low computing power conditions, and improving the efficiency, stability and relevance of scattering estimation. 5) Improved image restoration stability in complex underwater scenes: By introducing adaptive scattering compensation weights and restoration confidence weights, this invention can adjust the compensation intensity according to the scattering degree of different regions and suppress false restoration in low confidence regions, thereby reducing the risk of artifacts, over-enhancement and local distortion. 6) More conducive to recovering effective information obscured by scattering: The goal of this invention is not simply to improve the appearance of the image, but to prioritize the estimation of scattering and the recovery of target information obscured by scattering, thus better meeting the actual needs of underwater target observation, detection and engineering applications.

[0054] In summary, this invention provides a novel underwater image restoration technology solution suitable for low-computing-power embedded platforms. Under low-computing-power embedded platform conditions, instead of employing a full-image end-to-end enhancement approach, it concentrates the limited computing power of a lightweight neural network on the adaptive solution of underwater scattering components. That is, under low-computing-power conditions, it prioritizes solving the key problem of scattering, and then improves target visibility and restoration stability in scatter-weakened areas through scattering compensation, physical constraint restoration, and restoration confidence weight correction. Compared with existing underwater image enhancement, ordinary dehazing, and lightweight end-to-end network solutions, this invention has better application value in terms of restoration effect, stability, interpretability, and engineering feasibility.

[0055] Finally, it should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0056] The above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them; although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications can still be made to the specific implementation of the present invention or equivalent substitutions can be made to some technical features without departing from the spirit of the technical solutions of the present invention, and all such modifications and substitutions should be covered within the scope of the technical solutions claimed in the present invention.

Claims

1. A method for underwater image restoration for low-computing-power embedded platforms, characterized in that, Includes the following steps: S1. Acquire raw underwater degradation images , Indicates the pixel position in the image; S2, to Preprocessing is performed to obtain a standardized image. ;from Extract the fundamental physical features and denote their set as follows: ; S3, from Extract scattering-related features and construct a scattering indicator feature set. ; S4. Estimating scattering components using a lightweight neural network. , expressed as equation (8); where, This represents a lightweight neural network. Represents network parameters, Indicates network input characteristics, ;right Normalization is performed to obtain normalized scattering intensity. ; (8)； S5. Based on the estimated scattering components Standardized images Perform scattering compensation to obtain a scattering-compensated image. ; S6. Recover the target's visibility information weakened by scattering based on physical constraints to obtain the recovered image. , expressed as equation (14); where, This represents the estimated background light component. This represents the estimated transmittance. Indicates a stable term; (14)； S7, Based on scattering indicator feature set and normalized scattering intensity Obtain the recovery confidence weight ; and then according to equation (20) Make corrections to obtain the corrected restored image. ; (20)； S8, to Post-processing is performed to output the final restored image. .

2. The underwater image restoration method for low-computing-power embedded platforms according to claim 1, characterized in that, In step S3, a scattering indicator feature set is constructed according to equation (1). ;in, This represents the scattering indicator feature construction function. , , , , They represent Local average brightness, local contrast, color channel attenuation differences, edge response, and local blur degree; (1)。 3. The underwater image restoration method for low-computing-power embedded platforms according to claim 2, characterized in that, In step S4, the lightweight neural network Employs a lightweight encoder-decoder architecture. It consists of an input convolutional layer, several depthwise separable convolutional layers, a lightweight residual block, and an output convolutional layer. The input convolutional layer performs channel mapping; the depthwise separable convolutional layers extract local scattering features and reduce computational cost; the lightweight residual block enhances feature representation and preserves details; and the output convolutional layer outputs the estimated scattering components. .

4. The underwater image restoration method for low-computing-power embedded platforms according to claim 3, characterized in that, In step S5, the scattering compensation image is obtained according to equation (11). ,in, The scattering compensation weight is adaptively generated based on local scattering characteristics. (11)。 5. The underwater image restoration method for low-computing-power embedded platforms according to claim 4, characterized in that, In step S5, scattering compensation weights are generated according to equation (12). ;in, Indicates according to A function for adaptively calculating scattering compensation weights. This represents the truncation function. express Normalized local contrast , , For adjustment coefficients, and These represent the lower and upper limits of the scattering compensation weights, respectively. (12)。 6. The underwater image restoration method for low-computing-power embedded platforms according to claim 5, characterized in that, In step S6, Based on scattering compensation image In pixel position local window centered The low-frequency brightness or color statistics within the area are estimated using methods such as local mean, weighted average, or local bright candidate region mean. It is the normalized scattering intensity and The normalized local contrast is combined with the truncation function for estimation.

7. The underwater image restoration method for low-computing-power embedded platforms according to claim 6, characterized in that, In step S7, the recovery confidence weights are constructed according to equation (21). ;in, Indicates according to and A function for calculating the recovery confidence weights. This represents the truncation function. express Normalized local contrast express Normalized edge response, , , Represented as weighting coefficients, ; (21)。 8. An underwater image restoration system for low-computing-power embedded platforms, characterized in that, The underwater image restoration method for low-computing-power embedded platforms, as described in any one of claims 1 to 7, comprises: Image acquisition module, used to acquire raw underwater degraded images; The preprocessing module performs preprocessing operations on the original underwater degraded image, including resizing, color space transformation, and normalization, to obtain a standardized image and extract a set of basic physical features. The scattering indicator feature construction module is used to extract scattering-related features from the normalized image and construct a set of scattering indicator features; A lightweight neural network scattering estimation module for estimating scattering components in standardized images under low computational power conditions; The physical constraint recovery module is used to compensate the normalized image based on the estimated scattering components to obtain a scattering compensation image, and then recover the target visibility information weakened by scattering based on physical constraints to obtain a recovery image. Finally, the recovery image is corrected using recovery confidence weights. The post-processing output module performs post-processing operations on the corrected restored image, including detail smoothing, color consistency correction, brightness range constraint, and output format conversion, to output the final restored image.

Images