Abstract:

In order to improve the performance of small target detection in infrared imaging and the ability of low altitude airspace supervision, an infrared small target detection network based on multi-scale attention feature enhancement fusion was proposed. Firstly, Resnet34 was used to extract the multi-scale features of infrared images. Secondly, the multi-scale spatial attention feature enhancement module(MFEM) was used to improve the ability of feature extraction. Then, in the step-by-step up sampling process, the dual channel attention feature fusion module(DFFM) was used to fuse the semantic information and detail information to better protect the characteristics of infrared small targets. Finally, taking the video sequence detection of ground/air infrared dim small aircraft target as an example, the real scene test was carried out by comparing with other methods. The results show that compared with existing methods, the proposed method improves the scores of intersection over union(IoU), F-measure and false negative rate(FNR), and can accurately locate the target and generate good segmentation results. The DFFM can simultaneously use multi-scale context information and spatial attention mechanism to highlight infiared small targets. The DFFM assigns weights to sets of different channel features, thereby obtaining the most appropriate feature map for feature fusion and improving the detection performance.

Key words: infrared image, small target detection, feature enhancement, feature fusion, attention mechanism