DPFD-Net: a dual-path focus diffusion network for infrared small target detection

Infrared weak target detection is susceptible to low gray level difference, complex background, edge blur and other factors, resulting in sparse effective features and difficult separation of target and background., The mainstream frame is easy to introduce background noise due to the fixed receptive field, and it is difficult to maintain weak target features. In order to solve the above problems, this article constructs a dual-path focused diffusion network (DPFD-Net) based on Real-Time Detection Transformer (RT-DETR), and proposes an end-to-end infrared weak target detection framework, which can achieve fast inference on specified hardware. The framework is composed of a dual-path hybrid backbone network (DPHB) and a focused diffusion pyramid network (FDPN). The former combines dynamic convolution mixing and dynamic alignment mechanism to make up for the adaptability defect of single receptive field under weak infrared gradient, strengthen the feature expression of small and weak targets, and improve the robustness to labeling error and positioning offset. The latter uses focus-diffusion two-stage processing, which can effectively suppress background interference and enhance the ability of multi-scale feature discrimination by compressing the receptive field to highlight the weak target response and expanding the receptive field to fuse the global context. The test results on the A High-Altitude Infrared Thermal Dataset for Unmanned Aerial Vehicle (HIT-UAV) and Multi-Modal Multi-spectral Feature (M3FD-i) datasets show that compared with the baseline model, the average accuracy of PFD-Net is improved by 2.3% and 4.0%, respectively, and 6.7% and 5.2% improvements are achieved in the detection of small targets. At the same time, the number of model parameters is reduced from 19.9 M to 14.9 M. The experimental results fully prove the effectiveness and robustness of DPFD-Net for infrared small target detection in complex environments.