MobileViT Based Lightweight Model for Prohibited Item Detection in X-Ray Images

Abstract

The application of computer vision technology to detect prohibited items in X-ray security inspection images holds significant practical research value. We have observed that object detection models built upon the Visual Transformer (ViT) architecture outperform those relying on Convolutional Neural Networks (CNNs) when assessed on publicly available datasets. However, the ViT’s attention mechanism, while offering a global response, lacks the CNN model’s inductive bias, which can hinder its performance, demanding more samples and learning parameters. This drawback is particularly problematic for time-sensitive processes like security inspections. This research paper aims to develop a lightweight prohibited item detection model grounded in the ViT framework, utilizing MobileViT as the underlying network for feature extraction. To enhance the model’s sensitivity to small object features, we have established dense connections among various network layers. This design ensures effective integration of both high- and low-level visual features without increasing computational complexity. Additionally, learnable group convolutions are employed to replace traditional convolutions, further reducing model parameters and computational demands. Simulation experiments conducted on the publicly available SIXray dataset validate the effectiveness of the proposed model in this study. The code is publicly accessible at https://github.com/zhg-SZPT/MVray.

This work was supported by the School-level Project of Shenzhen Polytechinc University (No. 6022310006K), Research Projects of Department of Education of Guangdong Province (No. 2020ZDZX3082, 2023ZDZX1081, 2023KCXTD077).