In recent years, the Multi-layer convolutional sparse coding (ML-CSC) model has been regarded as a theoretical explanation for convolutional neural networks (CNN). While the ML-CSC model performs well on datasets with high feature contrast, its performance is not satisfactory on datasets with low feature contrast. To address this issue, this paper introduces a multi-scale technique to design a multi-scale multi-layer convolutional sparse coding network (MSMCSCNet), which not only achieves better image classification results in scenarios with weak feature contrast, but also provides the model with a solid theoretical foundation and higher interpretability. Experimental results demonstrate that, without increasing the parameter count, MSMCSCNet achieves accuracy improvements of 5.75, 9.75, and 9.8 percentage points on the Cifar10, Cifar100 datasets, and the Imagenet32 subset, respectively, compared to existing ML-CSC models. Furthermore, ablation experiments further validate the effectiveness of the model's multi-scale design and feature selection mechanism.

The deep learning technology has significantly improved the speed and accuracy of compressed sensing reconstruction. However, the existing deep compressive sensing algorithms usually use the same sampling rate to process different blocks of an image, ignoring the fact that different image blocks have different reconstruction difficulties. In this paper, a compressive sensing algorithm with adaptive sampling and memory enhancement is proposed. Firstly, the reconstruction difficulty of different blocks is estimated based on the reconstruction error of the measurement domain. Then, the rules are designed to adaptively assign the sampling rates, and the sampling matrix is used to sample each image block at a specific sampling rate with the help of a sampling rate mask. Furthermore, the two-branch aggregation module is added to the reconstruction network to enhance the interaction of context memory, and the reconstruction ability of the network is improved by adjusting the channel weight of different memory branches. The experimental results show that the proposed algorithm increases the average SSIM by approximately 0.0269 and the average PSNR by approximately 1.66 dB over other methods on several common datasets.