Image Compressed Sensing Based on Vision-Inspired Importance Maps

This paper presents a novel image compressed sensing framework that leverages human vision-inspired importance maps to guide the compression process. Unlike traditional methods that compress images without considering perceptual relevance, our approach selectively samples pixels based on their visual saliency and geometric significance. The framework begins by generating an importance map that combines visual saliency, obtained through LayerCAM, with geometric features like edge information. These maps dictate the adaptive sampling of pixels, ensuring that the most visually important information is retained. In the decoding stage, the sampled pixels are reconstructed using prior information-based models such as the Deep Image Prior and Manifold Modeling in Embedded Space. These models effectively restore the image while preserving its critical visual characteristics. To evaluate the performance of our framework, we introduce two new similarity criteria: Weighted PSNR (WPSNR) and Weighted SSIM (WSSIM), both tailored to assess the perceptual quality of images. Our experimental results demonstrate that the proposed method outperforms conventional image compression techniques, especially in maintaining perceptual quality at low sampling rates. This framework shows great promise for applications in areas where preserving visual fidelity is crucial, such as high-quality image compression and digital brain modeling. The integration of visual importance into the compression process represents a significant advancement in the field of image processing, offering a more efficient and perceptually aware approach to image compression.