$\hbox {ABDF}^{2}$-Net: an adaptive bi-directional features fusion network for real-time detection of threat object

Yiru Wei, Zhiliang Zhu, Hai Yu, Wei Zhang

Abstract: To achieve automatic detection of threat objects for X-ray baggage screening, we propose an adaptive bi-directional features fusion network (\(\hbox {ABDF}^{2}\)-Net) to detect threat objects on X-ray images. In \(\hbox {ABDF}^{2}\)-Net, an adaptive bi-directional feature fusion module (\(\hbox {ABDF}^{2}\hbox {M}\)) is introduced to fuse the multi-scale features from two directions, and the adaptive function is used to control the features passing rate. Besides, an atrous convolutional pyramid pooling (ACPP) is employed to capture global contextual information, which can provide global semantic guidance for multi-scale features. Finally, the fused multi-scale features are used to predict the final detection results through prediction modules. Experiments on the GDXray database demonstrate the effectiveness and superiority of our proposed method against the other four object detection methods.