Security X-ray image segmentation based on dense high-resolution parallel networks
To tackle the technical challenges posed by the diverse scale,disorderly arrangement,and overlapping occlusion of prohibited items in X-ray image detection for security checks,by improving the high-resolution network(HRNet)model and incorporating a de-occlusion unit,a novel multi-scale feature fusion network structure was proposed for prohibited item semantic segmentation in security X-ray images.In the encoding stage,leveraging HRNet's multi-branch parallel network structure,a dense connection was introduced within a single branch to enhance the fusion of information from deep and shallow layers,extracting multi-scale features to address the diverse scale of prohibited items in security X-ray images.Within the overall network architecture,an attention-based de-occlusion unit was integrated to enhance the model's edge perception,effectively suppressing the impact of overlapping occlusion on segmentation accuracy in security X-ray images.Experimental validation on the PIDray security image dataset across Easy,Hard and Hidden validation subsets demonstrated the effectiveness of the proposed method.The results indicate mean intersection over union(MIoU)values of 74.69％,69.92％and 56.77％,respectively.Compared with the original HRNet model,this represents an improvement of 2.03,1.62 and 4.13 percentage points,achieving an overall MIoU enhancement by approximately 2.59 percentage points.
万方数据
维普
