KECS-Net: Knowledge-Embedded CSwin-UNet With Slicing-Aided Hypersegmentation for Infrared Small Target Detection
Abstract
Infrared small target detection (IRSTD) remains a long challenging problem in infrared imaging technology. To enhance detection performance while more effectively exploiting target-specific characteristics, a novel U-shaped segmentation network called KECS-Net is proposed in this letter. KECS-Net first incorporates a CSwin Transformer module into the encoder of the UNet backbone, enabling the extraction of multi-scale features from infrared targets within an expanded receptive field, while achieving higher computational efficiency compared to the original Swin Transformer. Besides, a multi-scale local contrast enhancement module (MLCEM) is introduced, which utilizes hand-crafted dilated convolution operators to amplify locally salient target responses and suppress background noise, thereby guiding the model to focus on potential target regions. Finally, a slicing-aided hyper-segmentation (SAHS) method is also designed to resize and rescale the output image, increasing the relative size of small targets and improving segmentation accuracy during inference. Extensive experiments on three benchmark datasets demonstrate that the proposed KECS-Net outperforms state-of-the-art (SOTA) methods in both quantitative metrics and visual quality. Relevant code will be available at https://github.com/Lilingxiao-image/KECS-Net.