A novel underwater image restoration method based on decomposition network and physical imaging model

Abstract

Underwater image restoration is one of the significant research in marine engineering and aquatic robotics. However, due to the propagation characteristics of light and the serious turbidity in underwater, the captured images often have chromatic aberration and scattering blur, which brings great challenges to the restoration of the raw image. In this paper, a revised underwater imaging model is proposed first, which reanalyzes the generation of background light from the atmosphere to the underwater and provides important support for underwater color correction. And then a network framework via the revised model is designed, which can decompose the captured image into different components corresponding to the revised model. The proposed network consists of a decomposition architecture with residual blocks that learns a complete separation of clear image and transmittance features. These two features are used along with the raw image to predict the background light. Finally, combining three constraints of the imaging model, the proposed framework can converge rapidly along the desired direction. By comparison with the performance of the state‐of‐the‐art algorithms, the designed network shows excellent visibility and is capable of removing water on both synthetic and real‐world images in different water types.

A Two-Stage Screw Detection Framework for Automatic Disassembly Using a Reflection Feature Regression Model

“…In reflection removal, reflection features are regarded as noises that damage object characteristics. Here, researchers proposed various reflection removal methods mainly based on retinex theory, which assumes that an image can be decomposed into reflection and illumination components [ 22 , 23 , 24 ]. In reflection detection, reflection features are utilised as intrinsic features.…”

Privacy‐preserving image retrieval in a distributed environment

“…The global parameter for f L in the process of matrix encoding (5) and the public information P, on output the recovered similar images SI SI SI SI = { , , …, } l 1 2…”

“…Despite the tremendous benefits, image security and privacy become the most significant concern about cloud computing that must be well-addressed. 2,3 Image encryption [4][5][6] is a mature technology to protect image privacy from leaking in the cloud environment since it can completely hide the information of the original images. However, it impedes further image processing operations such as image retrieval.…”

An effective and practical gradient inversion attack

“…Through mathematical methods, we extract more information of the original pictures from the fully connected layer of the prior model, making the global optimization process closer to the ground‐truth image. Moreover, in image reconstruction, more consideration of network architecture is an effective method 26 . Therefore, inspired by the interpretability of CNN, 27 we innovatively proposed a denoising method according to the characteristics of CNN, denoising the characteristic image of each layer, so that the image can be recovered better in the iterative recovery process of the original method.…”

“…Moreover, in image reconstruction, more consideration of network architecture is an effective method. 26 Therefore, inspired by the interpretability of CNN, 27 we innovatively proposed a denoising method according to the characteristics of CNN, denoising the characteristic image of each layer, so that the image can be recovered better in the iterative recovery process of the original method.…”