关键词: 深度学习, 小目标检测, 注意力U型网络, 中心差分卷积, 对比学习

Abstract: Infrared Small Target Detection (IRSTD) is the most active research and suffers from difficulty in the field of target detection. IRSTD is difficult to learn accurate feature representation from few and deformable information of small target due to its characteristic of few pixels, low contrast and lacking texture, thus the IRSTD still faces the challenge. To address above issue, a multi-attention contrastive learning based IRSTD method was proposed. Firstly, the U-Net was adopted as the fundamental framework, a context mixed block (CMB) that integrates frequency attention and spatial attention was proposed during the encoding phase to produce a preliminary attention feature map. Then, in the decoding phase, a Multi-Kernel Central Difference Convolution (MKCDC) was devised to extract the core information of small targets, which remained stable across various scales. Finally, by combining binary cross-entropy loss and contrastive loss function for training small target detection network, the feature representation ability of small targets was enhanced and a discriminative small target detection model was achieved. The Probability of detection (Pd) of the proposed method on IRSTD dataset of 1000 infrared images (called IRSTD-1k) and Single-frame Infrared Small Target dataset collected by Nanjing University of Aeronautics and Astronautics (called NUAA-SIRST) reached 96.63% and 100.00% respectively, which were improved by 4.71% and 1.90% compared with DNANet (Dense Nested Attention Network), respectively. The experimental results show that the proposed method effectively improves the performance of infrared small target detection.

Key words: deep learning, infrared small target detection, attention U-Net, central difference convolution, contrastive learning