面向安检X光图像的违禁物品语义分割与识别研究

针对安检X光图像中违禁物品大小不一、物品摆放随意且存在重叠遮挡的技术难题，提出了一种改进的HRNet多尺度特征融合网络模型，实现图像中违禁物品的自动分割与识别。在编码阶段，利用HRNet网络中的多分辨率并行网络架构，提取多尺度特征，解决安检X光图像违禁物品尺度多样化的问题。在解码阶段，提出一种多层级特征聚合模块，采用数据相关上采样方法减少信息丢失，并聚合编码阶段提取的特征，以对物品进行更完整表征。在网络整体架构中，嵌入基于注意力机制的去遮挡模块加强模型的边缘感知能力，缓解安检X光图像中物品重叠遮挡严重的问题，提高模型的分割识别精度。通过在PIDray安检图像公开数据集进行实验，结果表明，在Easy、Hard、Hidden 3个验证子集上分别取得了73.15%、69.47%、58.33%的平均交并比，相比原始HRNet模型，分别提升了0.49%、1.17%、5.69%,总体平均交并比提升约2.45%。

Semantic segmentation and recognition of contraband for security X-ray images

In response to the technical challenges posed by the varying sizes, haphazard arrangement, and overlapping occlusion of prohibited items in security X-ray images, we propose an enhanced HRNet-based multi-scale feature fusion network model. This model aims to achieve automatic segmentation and recognition of prohibited items in images. In the encoding stage, we leverage the multi-resolution parallel network architecture of HRNet to extract multi-scale features, addressing the diverse scale of prohibited items in security X-ray images. In the decoding stage, a multi-level feature aggregation module is introduced that uses data-dependent upsampling instead of bilinear interpolation. upsampling to reduce information loss during aggregation, thus ensuring a more comprehensive representation of the features of the features extracted in the coding stage for a more complete characterisation of objects.In the overall architecture of the network, a de-obscuration module based on the attention mechanism is embedded to strengthen the edge-awareness ability of the model, alleviate the problem of serious overlapping occlusion of items in security X-ray images, and improve the segmentation and recognition accuracy of the model. By experimenting on the public dataset of PIDray security check images, the results show that the average intersection ratio of 73.15%, 69.47%, and 58.33% are achieved in the three validation subsets of Easy, Hard, and Hidden, respectively, which are 0.49%, 1.17%, and 5.69%, respectively, and the overall average intersection ratio is improved by about 2.45%.