难点注意力感知红外小目标检测网络OA北大核心CSTPCD

As aircraft maneuverability increases,multi-frame infrared small target detection methods are be-coming insufficient to meet detection requirements.In recent years,significant progress has been achieved in single-frame infrared small-target detection method based on deep learning.However,infrared small targets often lack shape features and have blurred boundaries and backgrounds,obstructing accurate segmentation.According to the problems,an indistinguishable points attention-aware network for infrared small object de-tection was proposed.First,potential target areas were acquired through a point-based region proposal mod-ule while filtering out redundant backgrounds.Then,to achieve high-quality segmentation,the mask bound-ary refinement module was utilized to identify disordered,non-local indistinguishable points in the coarse mask.Multi-scale features of these difficult points were then fused to perform pixel-wise attention modeling.Finally,A fine segmentation mask was generated through re-predicting the indistinguishable points attention-aware features by point detection head.The mAP of the proposed method reached 87.4 and 63.4 on the pub-licly available datasets NUDT-SIRST and IRDST,and the F-measure reached 0.8935 and 0.7056,respect-ively.It can achieve accurate segmentation in multi-detection scenarios and multi-target morphology,sup-pressing false alarm information while controlling the computational overhead.