Search

CN-122023158-A - Underwater low-light image recovery method and device, electronic equipment and storage medium

CN122023158ACN 122023158 ACN122023158 ACN 122023158ACN-122023158-A

Abstract

The invention relates to an underwater low-light image recovery method, an underwater low-light image recovery device, electronic equipment and a storage medium, wherein the method comprises the steps of acquiring and processing an original underwater polarized image and calculating an orthogonal polarized image; three parallel depth image prior networks are constructed and initialized and are respectively used for generating direct transmitted light and orthogonal polarized state back scattered light, the networks are combined and optimized under the polarization physical constraint, the accurate decoupling and illumination recovery of scattering components are realized by minimizing the total loss function formed by reconstruction loss, polarization difference loss, polarization degree reconstruction loss and self-adaptive stretching loss, and the direct transmitted light and the back scattered light generated by the optimized networks are used for recovering a final underwater clear image according to an underwater imaging model. The invention effectively integrates the physical imaging model and the data driving advantages, and realizes the recovery of the underwater image with robustness and high quality aiming at the complex degradation of the underwater low-light scene.

Inventors

  • HU HAOFENG
  • XU YIJIA
  • FEI XIAOTONG
  • SHEN LINGHAO

Assignees

  • 天津大学

Dates

Publication Date
20260512
Application Date
20260128

Claims (8)

  1. 1. The method for recovering the unsupervised underwater low-light image based on polarization guidance is characterized by comprising the following steps of: s1, acquiring and processing an original underwater polarized image, and calculating an orthogonal polarized image; S2, constructing and initializing three parallel depth image prior networks which are respectively used for generating direct transmission light and orthogonal polarization state back scattering light; s3, the network is jointly optimized under the constraint of polarization physics, and the accurate decoupling and illumination recovery of the scattering component are realized by minimizing a total loss function consisting of reconstruction loss, polarization difference loss, polarization degree reconstruction loss and self-adaptive stretching loss; s4, recovering a final underwater clear image according to the underwater imaging model by utilizing the direct transmitted light and the back scattered light generated by the optimized network.
  2. 2. The polarization-guided unsupervised underwater low-light image restoration method according to claim 1, wherein in step S1, the images of four polarization angles of 0 °, 45 °, 90 ° and 135 ° are obtained by single exposure of the split focal plane polarization camera , , , Calculation of Stokes vector , , Thereby obtaining a total light intensity image And images of orthogonal polarization, i.e. parallel polarization And vertical polarization state image The calculation formula is as follows: ; ; Wherein, the , , 。
  3. 3. The method for recovering an unsupervised underwater low-light image based on polarization guidance according to claim 1, wherein in step S2, three parallel DIP networks are constructed as follows: 、 、 Wherein, the method comprises the steps of, And All take grid noise as input for generating back scattered light with parallel and vertical polarization states And ; The network takes random noise as input for generating direct transmission light ; Wherein, the 、 And Respectively networks 、 、 The corresponding input noise is used to determine the output noise, For the direct transmitted light generated by the network, And For network-generated back-scattered light of orthogonal polarization, and = ; The network structure is designed differently for different physical roles, namely, a direct transmission optical network With five layers of codecs, 128 channels per layer, using a 3 x 3 convolution kernel to preserve detail, a back-scattered optical network And With a lightweight design, the number of channels is [8,16,32,64,128], and a 5×5 convolution kernel front three-layer jump-free connection is used to reduce high-frequency information transmission.
  4. 4. The polarization-guided-based unsupervised underwater low-light image restoration method according to claim 1, wherein in step S3, the total loss function L of network optimization consists of four parts: Wherein, the 、 、 And The loss weights are 1,5 and 5 respectively; The function and design of each loss term is as follows: Reconstruction loss Ensuring that scattered light and transmitted light components output by a network can accurately reconstruct an original orthogonal polarized image, wherein the method comprises the following steps: polarization differential constraint of back-scattered light The method comprises the steps of directly restricting the difference between the orthogonal polarization state back scattered light to be consistent with the polarization difference of an original image by utilizing polarization difference information contained in the original image, guiding a network to learn the correct polarization characteristic of the scattered light, and representing the following steps: ; polarization degree reconstruction constraints Requiring the degree of polarization of the image reconstructed by the network Degree of polarization with original image Keeping consistent, providing constraints on overall polarization properties, expressed as: wherein the polarization degree information is reconstructed And original polarization degree information The following is shown: Adaptive loss Aiming at the problem of low contrast of low-light images, for vertical polarization state images Performing self-adaptive stretching based on pixel statistics to obtain And restrict direct transmitted light to approach, effectively enhancing image contrast and texture, expressed as: Wherein, the And Respectively is Pixel values of 1% and 99% for determining the adaptive stretch range; Three networks using Adam optimizer 、 And And performing joint optimization, iterating 5000 times, wherein the learning rate is 0.001, and adding Gaussian disturbance with standard deviation of (1/30) to network input noise as a regularization strategy in each iteration.
  5. 5. The method for recovering an unsupervised underwater low-light image based on polarization guidance according to claim 1, wherein in step S4, the reflected light of the real object is recovered according to the underwater polarization imaging model The expression is as follows: Wherein, the For the direct transmitted light generated by the network, t is the transmittance, expressed as follows: back-scattered light generated from a network Substituting the formula can recover the reflected light of the actual object 。
  6. 6. An underwater low light image restoration device for implementing the polarization guide-based unsupervised underwater low light image restoration method according to any one of claims 1 to 5, comprising: The polarized image processing unit is used for acquiring an original underwater polarized image through the polarized camera, calculating a Stokes vector and an orthogonal polarized image, and providing polarized image data for a subsequent unit; the network construction and initialization unit is used for constructing three parallel depth image priori networks with different structures, inputting random noise and grid noise respectively to initialize network parameters, and providing an initialization network for the subsequent units; the physical constraint optimization unit is connected with the polarized image processing unit and the network construction and initialization unit and is used for jointly optimizing the network under the physical constraint formed by the reconstruction loss, the polarization difference loss, the polarization degree reconstruction loss and the self-adaptive loss so as to realize decoupling and illumination recovery of scattered light and target light; the image recovery unit is connected with the polarized image processing unit and the physical constraint optimizing unit and is used for calculating and outputting a final underwater clear image by utilizing the direct transmitted light and the back scattered light generated by the optimized network and combining an underwater polarized imaging physical model.
  7. 7. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the polarization-guided unsupervised underwater low-light image restoration method according to any of claims 1 to 5 when the program is executed by the processor.
  8. 8. A computer readable storage medium, having stored thereon a computer program, which when executed by a processor, implements the polarization-guided based unsupervised underwater low-light image restoration method according to any of claims 1 to 5.

Description

Underwater low-light image recovery method and device, electronic equipment and storage medium Technical Field The present invention relates to the field of underwater image processing technologies, and in particular, to an unsupervised underwater low-light image restoration method and apparatus based on polarization guidance, an electronic device, and a computer readable storage medium. Background The underwater image is seriously influenced by the scattering and absorption effects of the water body in the shooting process, so that degradation problems such as contrast reduction, color distortion, detail blurring and the like of the image captured by the detector are generally caused. Under low light imaging conditions such as deep sea or night, significant attenuation of illumination further exacerbates degradation of image quality, making effective restoration very challenging. Aiming at the problems, the prior art scheme mainly can be divided into two main types of a polarization imaging restoration method based on a physical model and a deep learning enhancement method based on data driving: 1. the method uses the difference of the scattered light and the target reflected light in polarization characteristics to perform imaging and restoration. However, in low light environments, the signal-to-noise ratio drops dramatically, and the prior assumptions of the atmosphere or water on which conventional methods rely often no longer hold, resulting in misalignment of the estimation of critical optical parameters (e.g., transmittance, background light). This not only affects the effectiveness of scattered light removal, but also tends to amplify image noise during the restoration process, resulting in poor visual results. 2. The data driving method learns the end-to-end mapping from the degraded image to the clear image through a training network. However, the dominant supervised learning paradigm relies heavily on large scale, precisely paired training data sets (i.e., sharp-degenerate image pairs). In an actual underwater low-light scene, the cost for acquiring the real and paired data is extremely high and difficult, and the serious data scarcity problem restricts the training and generalization capability and the actual application performance of the model. In addition, some existing unsupervised or weakly supervised methods, while reducing reliance on paired data, often lack explicit modeling and constraints on underwater imaging physics processes. When the method is applied to a complex low-light environment with strong coupling of scattering and noise, a network is difficult to accurately decouple a scattering component and a target signal from a single Zhang Tuihua image, so that problems such as detail loss, color deviation or incomplete decoupling exist in a restoration result. Therefore, how to effectively combine the reliability of the physical model and the flexibility of data driving without depending on paired training data, so as to realize Lu Bangjie coupling of scattered light and high-quality image restoration in underwater low light field scenes, has become a technical problem to be solved in the field. Disclosure of Invention The invention aims to solve the problems of prior failure, data dependence, inaccurate decoupling and the like in the underwater low-light image restoration in the prior art, and provides an unsupervised underwater low-light image restoration method, device, electronic equipment and computer readable storage medium based on polarization guidance. The invention adopts the following technical scheme to realize the aim: an unsupervised underwater low-light image recovery method based on polarization guidance comprises the following steps: s1, acquiring and processing an original underwater polarized image, and calculating an orthogonal polarized image; S2, constructing and initializing three parallel depth image prior networks which are respectively used for generating direct transmission light and orthogonal polarization state back scattering light; s3, the network is jointly optimized under the constraint of polarization physics, and the accurate decoupling and illumination recovery of the scattering component are realized by minimizing a total loss function consisting of reconstruction loss, polarization difference loss, polarization degree reconstruction loss and self-adaptive stretching loss; s4, recovering a final underwater clear image according to the underwater imaging model by utilizing the direct transmitted light and the back scattered light generated by the optimized network. In particular, in step S1, images of four polarization angles of 0 °, 45 °, 90 °, 135 ° are acquired by a single exposure with the split focal plane polarization camera,,,Calculation of Stokes vector,,Thereby obtaining a total light intensity imageAnd images of orthogonal polarization, i.e. parallel polarizationAnd vertical polarization state imageThe calculation formula is as follows: