Search

Infrared small target detection has been widely used in the military and civil fields. The detection difficulties lie in small target size, few details and complex background interference. Existing classic deep learning detection methods are commonly suitable for generic target detection, but have poor adaptability to infrared small targets. To handle these problems, this article build a new infrared small target detection network based on U-shaped attention block and pixel-wise attention block. Firstly, a residual U-shaped attention block to extract multi-scale features through local U-shaped subnetworks in a single layer level and extract multiscale features to enrich the representation of small-scale target features is designed, so as to enhance the network′s ability to discriminate small-scale targets. Then, small target information is further preserved through dense fusion method to alleviate the semantic gap feature fusion between different layers and reduce the miss detection rate. Finally, the pixel-wise attention in space and channel dimensions is applied to the fused feature map to enhance small targets and suppress complex background interference. The experimental results show that the present network outperforms the latest benchmark method in the intersection over union, detection probability and false alarm rate of two infrared small target data sets NUDT-SIRST and IRSTD-1k. Moreover, the present network also achieves a good balance between the detection accuracy and the efficiency.