Effective Infrared Small Target Detection Based on Improved RefineDet

Aiming to improve the accuracy and efficiency of infrared (IR) small target detection, an end-to-end multi-scale improved RefineDet method is proposed. There are mainly two contributions. First, based on the two-step cascade of RefineDet, a small anchor refinement module (SARM) to extract features with strong compactness and selectivity of the human visual system is proposed. SARM use dilated convolution to expand the receptive field and add convolution branches to simulate the human visual system. Second, an improved feature attention transfer connection block (FA-TCB) is proposed. The FA-TCB transmits more helpful features by adding a channel attention mechanism in TCB to accumulate features recalibration. In addition, an IR high-altitude small target dataset is constructed, including 1200 images belonging to three categories, i.e., aeroplanes, UAVs, and birds. Experimental results show that the proposed strategies are beneficial for improving the performance of IR small target detection. By extracting features with strong compactness and paying more attention to channel dependency, the proposed method achieves superior performance than other well-known detection methods.