Picture segmentation is a crucial yet challenging problem in the domains of computer vision and image processing. Applications that use it include medical image processing, content-based image retrieval, target tracking, and object identification. The goal of image segmentation is to divide an image into a predetermined number of sections that have the same color, texture, and other characteristics. We then cluster the relevant areas together for a more legible graphic. At the low level, splitting an image into two “simpler” layers has been a common way to do things like picture recovery and enhancement. There are more unknowns than inputs, making the problem ill-posed. A task-aware prior estimate and a decomposition model are the two parts of the method that this article talks about. It is suggested that a pixel-wise analysis sparsity model can be used to make the separation layers more even in the analysis operator's assumed sparse converted image. Instead of regularizing all pixels with a single penalty weight, we try to achieve pixel-wise sparse penalties by estimating each pixel's sparsity level using task-aware priors. Furthermore, to take advantage of their complementary processes, one separation layer is regularized using both the synthesis sparsity model and the pixel-wise analysis sparsity model. Unlike the analysis approach that employs image local features, the synthesized method uses non-local similarity cues and an over-complete dictionary to provide a flexible prior for regularizing the decomposition results. We use an alternating optimization technique to solve the proposed model. We evaluate it using the application for removing rain streaks and the Retinex model. Comprehensive experiments on many enhancement datasets, a large number of synthetic images, and real rainy images demonstrate that our method can remove imaging noise during Retinex decomposition and provide high-fidelity detraining results.