Multi-Scale X-ray Semantic Segmentation for Security Inspection Using Transformer with Haar wavelet Downsampling

To address the challenge of contraband segmentation in X-ray security images, this paper proposes a Transformer-based X-ray image segmentation network, HTnet, which incorporates Haar wavelet downsampling.First, a Transformer algorithm based on Haar wavelet Downsampling (HWD) is introduced to improve segmentation performance for X-ray images.The Haar wavelet downsampling operation effectively reduces the spatial resolution of feature maps while preserving edge and texture details, thereby retaining high-quality contraband contours and reducing uncertainty in the information.Secondly, to extract spatial details and multiscale hidden information more effectively, a Deformable Convolutional Pyramid Structure (DCPS) is proposed, which dynamically adjusts to accommodate varying contraband shapes.Finally, the Convolutional Block Attention Module (CBAM) is incorporated to enhance the network's spatial and channel feature learning, guiding it to focus more accurately on contraband regions within X-ray images.Experiments on public X-ray contraband datasets demonstrate that HTnet achieves superior performance in terms of the mIoU metric, significantly outperforming other state-of-the-art segmentation algorithms.