Interpretable Infrared Target Detection via Multi-Scale Robust Principle Component Analysis

Infrared small target detection faces challenges of minimal target size, low signal-to-noise ratios, and complex backgrounds. Deep learning methods achieve impressive results, which lack theoretical interpretability. Although the Robust Principle Component Analysis (RPCA) theory has been integrated with deep learning to build up RPCANet, this network is limited by single-scale decomposition. To this end, we formulate infrared small target detection task as a Multi-scale RPCA Optimization (MRO) model, which can be unfolded as the MRO-Net. Considering that the MRO network does not explore cross-stage interaction and image-domain weighted fusion is overemphasized, we extend the standard MRO-Net as MRO-Net+ by inserting multi-scale decomposition block from each stage of MRO model into the encoder, bottle-neck layer, and decoder of high-throughput dynamic U-shape architecture. Experimental results have shown the proposed explicable MRO-Net+ has better performances than several state-of-the-art methods such as RPCANet, UIUNet, DNANet, AGPCNet, etc.