Salient Region Guided Colour Image Restoration using Deep Learning with Adaptive Compressive Sensing

In this research, a unique colour image reconstruction method based on Compressive Sensing is presented, aimed at enhancing efficiency and quality. The proposed method incorporates deep learning techniques, adaptive measurement rate setting, and salient region detection. The process begins with the collection of an input RGB images, followed by pre-processing to convert them to the Y+ colour space., including optimized Faster R-CNN (OFRCNN), are employed to detect and segment salient regions in the Y-channel. An adaptive measurement rate setting using Deep Q-Network (DQN) dynamically adjusts measurement rates based on input data characteristics. To optimize the Y-channel compression, a Dual attention-based CS Network (CSNet) model is applied. Moreover, adaptive weighted reconstruction is introduced, utilizing salient region masks and a hybrid the Walrus Optimization Algorithm (WaOA) and mountain gazelle optimizer (MGO), called Hybrid Walrus with Mountain Gazelle Optimizer (HWMGO). Finally, the restored Y-channel is combined with the reconstructed U and V channels to reconstruct full-colour in the RGB colour space. Experimental evaluations on a diverse dataset demonstrate the effectiveness of the proposed algorithm in accurately reconstructing high-quality colour images from CS data. This research contributes to the advancement of colour image reconstruction techniques, offering a comprehensive and efficient framework for various applications.