Infrared small target detection is usually limited by a long imaging distance,which makes it difficult to extract target features.How to enhance target feature expression is one of the main research directions in recent years.However,too complex feature representation will lose the speed of inference.In this paper,we use reparameterization technology and residual network as feature enhancement module and feature fusion module,and achieve good results on the datasets.On SIRST and IRSTD-1K datasets,the proposed method achieves 0.734 and 0.638 mIoU,while having only 0.306M and 1.114G FLOPs in parameter number and computational complexity.Our model can maintain fewer parameters in the inference stage while having performance similar to or even leading other leading methods,which has obvious advantages in a serial environment.
关键词
红外小目标检测/深度学习/卷积神经网络/模型压缩/注意力机制
Key words
infrared small targets detection/deep learning/convolutional neural networks/model compression/attention mechanism
维普
万方数据