CN-122024031-A - Underwater image intelligent real-time enhancement method and device and electronic equipment

CN122024031ACN 122024031 ACN122024031 ACN 122024031ACN-122024031-A

Abstract

The invention relates to an intelligent real-time enhancement method and device for an underwater image and electronic equipment, wherein the method comprises the steps of obtaining the underwater image to be processed; the method comprises the steps of carrying out rapid degradation analysis on an underwater image, extracting a low-dimensional statistical feature set, comparing the low-dimensional statistical feature set with a preset threshold value set to determine degradation scene types of the underwater image, dynamically modulating internal processing parameters of a lightweight converter enhancement engine according to the degradation scene types, wherein the internal processing parameters comprise feature fusion weights in a multi-scale feature fusion decoder, carrying out enhancement processing on the underwater image by utilizing an enhancement engine after parameter modulation, and outputting the enhanced image. The invention can realize real-time processing of the high-resolution image under the constraint of limited computing force, and effectively identify and adaptively process various underwater image degradation scenes including composite degradation.

Inventors

ZHU DAQI
GUO JIAHAO
JI HAOMING
LI HONGFEI
CHEN MINGZHI

Assignees

上海理工大学

Dates

Publication Date: 20260512
Application Date: 20260414

Claims (10)

1. An intelligent real-time enhancement method for an underwater image, which is characterized by comprising the following steps: acquiring an underwater image to be processed; Carrying out rapid degradation analysis on the underwater image, and extracting a low-dimensional statistical feature set, wherein the low-dimensional statistical feature set consists of three statistical features of a whole frame gray histogram mean value, a local contrast space mean value and a red-blue channel gray difference value; Comparing the low-dimensional statistical feature set with a preset threshold value set, and determining a degradation scene type of the underwater image, wherein the degradation scene type at least comprises a low light-turbidity composite degradation scene, and the identification condition of the low light-turbidity composite degradation scene is that the average value of the whole frame gray level histogram is lower than a first threshold value and the spatial average value of the local contrast is lower than a second threshold value; Dynamically modulating internal processing parameters of a lightweight fransformer enhancement engine according to the determined degradation scene category, wherein the internal processing parameters comprise feature fusion weights in a multi-scale feature fusion decoder; the underwater image is enhanced by the light-weight transducer enhancement engine after parameter modulation, and the enhanced image is output; The lightweight converter enhancement engine comprises a local sparse attention module, a cross-window information interaction module and a multi-scale feature fusion decoder, wherein the local sparse attention module is used for dividing an input feature map into non-overlapping windows and independently carrying out attention calculation in each window, the cross-window information interaction module is used for introducing a global downsampling and upsampling path after the local sparse attention module to realize inter-window information transfer, and the multi-scale feature fusion decoder is used for fusing multi-level features extracted by an encoder and carrying out weighted fusion on detail features from shallow layers of the encoder and deep semantic features according to feature fusion weights dynamically generated by degradation scene types.
2. The method for intelligent real-time enhancement of an underwater image according to claim 1, wherein the calculation of the whole frame gray histogram mean value comprises the steps of converting the underwater image into a gray image, and calculating the arithmetic mean value of gray values of all pixels of the gray image; Dividing the underwater image into a plurality of local areas with equal size, calculating the local contrast of each local area, and calculating an arithmetic average value of the local contrast of all the local areas; The calculation of the gray scale difference value of the red and blue channels comprises the steps of respectively extracting a red channel and a blue channel of the underwater image, and calculating the difference between the average value of the gray scale values of all pixels of the red channel and the average value of the gray scale values of all pixels of the blue channel.
3. The method for intelligent real-time enhancement of an underwater image according to claim 1, wherein the preset threshold set is calibrated based on massive in-field deep sea image data, and the degraded scene category further comprises a severe low light scene, a moderate cloudy scene and a normal scene.
4. The method for intelligent real-time enhancement of an underwater image according to claim 3, wherein the identification condition of the severe low-light scene is that the whole frame gray histogram mean value is lower than a first threshold value and the local contrast spatial mean value is not lower than a second threshold value; The identification condition of the medium-turbidity scene is that the local contrast space average value is lower than a second threshold value and the whole frame gray histogram average value is not lower than a first threshold value; the identification condition of the low light-turbidity composite degradation scene is that the average value of the whole frame gray level histogram is lower than a first threshold value and the spatial average value of the local contrast is lower than a second threshold value; The identification condition of the normal scene is that the average value of the whole frame gray level histogram is not lower than a first threshold value and the spatial average value of the local contrast is not lower than a second threshold value.
5. The method of claim 1, wherein the local sparse attention module divides the input feature map into non-overlapping windows of size M x M, where M is an integer greater than 1, and the attention computation is performed only inside each window.
6. The method for intelligent real-time enhancement of underwater images according to claim 5, wherein the cross-window information interaction module adopts a shift window strategy to realize information transfer between windows.
7. The underwater image intelligent real-time enhancement method according to claim 5, wherein the cross-window information interaction module comprises a downsampling branch for downsampling the output feature map of the local sparse attention module to extract global context information, an upsampling branch for restoring the global context information to an original resolution, and a fusion unit for fusing the restored global context information with the output feature map of the local sparse attention module.
8. The intelligent real-time enhancement method for underwater images according to claim 1, wherein the feature fusion weights dynamically generated by the multi-scale feature fusion decoder adjust the fusion ratio of the detail features from the shallow layer and the semantic features of the deep layer of the encoder according to the degradation scene category, wherein when a high-turbidity scene is handled, higher weights are allocated to detail feature channels to promote edge sharpening and texture recovery, and when a low-light scene is handled, higher weights are allocated to semantic feature channels to emphasize global brightness stretching and noise suppression.
9. The method of claim 1, wherein dynamically modulating internal processing parameters of a lightweight transform enhancement engine according to the determined degradation scene category comprises: Aiming at a low-light scene, activating a low-light optimization branch, and injecting brightness priori mapping in a model input stage; Aiming at a high-turbidity scene, a turbidity removing branch is activated, and a model is guided to simulate the inverse process of a physical scattering process; aiming at a low light-turbidity composite degradation scene, a cooperative processing assembly line is started, and the basic brightness enhancement of low light enhancement is firstly executed, and then the turbidity removing operation is executed.
10. An electronic device comprising a memory and a processor, the memory storing a computer program, characterized in that the processor implements the underwater image intelligent real-time enhancement method according to any of claims 1 to 9 when executing the computer program.

Description

Underwater image intelligent real-time enhancement method and device and electronic equipment Technical Field The invention relates to the technical field of underwater image enhancement and visual perception, in particular to an intelligent real-time underwater image enhancement method and device and electronic equipment. Background When the deep sea submersible performs navigation, observation and operation tasks, the visual perception system of the deep sea submersible is highly relied on to acquire clear and reliable underwater images. However, the deep sea imaging environment is extremely complex, and the absorption and scattering effects of seawater on light cause various degradation problems of acquired images. Specifically, the rapid attenuation of illumination along with depth leads to low overall brightness of an image, loss of dark detail and remarkable noise (low illumination degradation), and the light scattering caused by suspended particles in a water body leads to the reduction of contrast of the image, and haze blurring and color distortion (turbidity degradation) appear. In practical operation, low light and high turbidity conditions often coexist to form a more complex composite degradation scene, which poses serious challenges to the usability of the vision system. In order to improve the quality of underwater images, image enhancement methods based on deep learning have been widely used. The visual transducer model has potential in image restoration tasks due to strong global context modeling capability. However, there is a significant technical bottleneck in applying it to the real-time image enhancement tasks faced by deep sea vehicles. First, the computational complexity of the self-attention mechanism in standard Transformer architecture is proportional to the square of the number of image pixels, and for high resolution underwater images (e.g., 1080 p), its huge computational overhead exceeds the real-time processing power of the on-board embedded edge computing platform of the submersible (e.g., JTXavier series), which typically requires processing delays below 33ms, resulting in a difficult reconciliation between model performance and real-time requirements. Secondly, most of the existing enhancement models are of fixed structures, and a mechanism for rapidly and accurately identifying specific degradation types (especially low light, turbidity and combination of the low light and the turbidity) of an input image is lacked, so that internal processing strategies cannot be dynamically adjusted according to actual degradation scenes. A general model trained on a mixed data set is often not clear enough in the enhancement direction when facing specific dominant degradation, and may cause poor processing effect in a single degradation scene, and multiple degradation factors are difficult to cooperatively process in a composite degradation scene, so that pertinence, adaptability and reliability of a final enhancement effect of the model are limited. Therefore, it is needed to design an underwater image enhancement scheme suitable for an edge computing environment, so that the underwater image enhancement scheme can realize real-time processing of high-resolution images under the constraint of limited computing force, and can effectively identify and adaptively process various underwater image degradation scenes including compound degradation so as to meet the severe requirements of a deep sea submersible visual perception system on high reliability, high instantaneity and strong environmental adaptability. Disclosure of Invention Based on the above, it is necessary to provide an intelligent real-time enhancement method, device and electronic equipment for underwater images, which can run on edge computing equipment in real time, have adaptive degradation scene recognition capability and have a targeted processing strategy for composite degradation scenes. The invention provides an intelligent real-time enhancement method for an underwater image, which comprises the following steps: acquiring an underwater image to be processed; Carrying out rapid degradation analysis on the underwater image, and extracting a low-dimensional statistical feature set, wherein the low-dimensional statistical feature set consists of three statistical features of a whole frame gray histogram mean value, a local contrast space mean value and a red-blue channel gray difference value; Comparing the low-dimensional statistical feature set with a preset threshold value set, and determining a degradation scene type of the underwater image, wherein the degradation scene type at least comprises a low light-turbidity composite degradation scene, and the identification condition of the low light-turbidity composite degradation scene is that the average value of the whole frame gray level histogram is lower than a first threshold value and the spatial average value of the local contrast is lower than a second threshol