PIXDet: Prohibited Items X-Ray Image Detection in Complex Background

Abstract

In this paper, aiming at the complex background and overlapping characteristics in X-ray images, we propose an unique spatial attention mechanism based on the feedback of high-level semantic feature to guide low-level semantic features, named Feedback Guidance Mechanism (FGM). In addition, in view of the high probability of miss of small prohibited items, a feature aggregation method based on the fusion of high and low-level features and dilated convolution is proposed, named Feature Aggregation Module (FAM). Then, we combine FGM and FAM into a lightweight model SSD and get a new Prohibited Items Detector (PIXDet). Our experiments indicate that PIXDet is more lightweight, but it can achieve 90.36% mAP on PIXray dataset, exceeding SSD by 1.0% mAP, outperforming some state-of-the-art methods, implying its potential applications in prohibited items detection field.

Attention‐based prohibited item detection in X‐ray images during security checking

“…During the high-layer semantic feedback model, 𝛼 l can be calculated by deconvoluting the output with the parameters of the convolutional layer. The bottom-up transmission of features can be represented by Equation ( 4), while the top-down attention response is calculated by Equation (5).…”

“…Under such high-intensity working environment, physical and mental fatigue can cause the accidents, such as missed, wrong and forgotten inspection etc. Intelligent measures for automatic prohibited items detection are therefore favorable for alleviating the working pressure of security personnel as well as improving the operating efficiency [4,5].…”