| Popis: |
Adaptive Block-Based Compressed Sensing (ABCS) enables optimization of image and video sensing platforms with limited resources, using novel algorithms for efficient reconstruction and real-time operations. Taking number of measurements adaptively based on information contained in each block of the image, results in better quality of recovered image, and is paving ways for general purpose use of compressed sensing in various applications. Some of the major challenges in ABCS are, complexity introduced at encoder end for adaptive rate allocation and choosing or learning a measurement matrix for each allocation rate. As part of this paper, we introduce a novel adaptive measurement allocation technique based on edge information with in the block. The algorithm is further improved by giving special considerations for edges present at the block boundary. DCT (Discrete Cosine Transform) dictionaries are used for measurement and recovery. Using simple encoding for measurement allocation greatly reduces the complexity of encoder network in this scheme. Recovery is done by simple IDCT (Inverse Discrete Cosine Transform) in case of DCT dictionaries. This approach demonstrates high effectiveness without the need for computationally expensive GPU-based training. To assess the network’s generalizability, we conducted tests using both natural and medical images. Remarkably, the method exhibited consistent accuracy across various measurement rates in both scenarios. The recovery time for compressively sensed images, whether they are natural or medical, is real-time, with an average duration of around 50-65 milliseconds. The algorithm is also used in conjunction with Content Aware Scalable deep compressed sensing Network (CASNET) to get learned matrix for measurements on encoder side and pretrained model for reconstruction on decoder side. Proposed method not only converges in constant time across blocks in contrary to the rate allocation method of CASNET, but also outperforms the recovery quality in lower measurement rates. Extensive experimental results shows that proposed algorithm out performs other state of the art algorithms recovery. |
