CFPA-Net: Cross-layer Feature Fusion And Parallel Attention Network For Detection And Classification of Prohibited Items in X-ray Baggage Images

As objects in the baggage are often heavily overlapped and cluttered, the X-ray baggage inspection is an inherently challenging task. In this paper, we propose a cross-layer feature fusion and parallel attention network named CFPA-Net to detect and classify the prohibited items in X-ray baggage images. The CFPA-Net is based on RetinaNet with three modules: cross-layer feature extraction fusion module (CEF-Module), paralleled attention module (PA-Module) and FreeAnchor. In CEF-Module, an improved feature pyramid network is proposed by adding multi-directional lateral connections for cross-layer feature extraction and fusion. It can help detect objects of various scales and supplement deficient semantic and localization information for low layer and high layer features respectively. PA-Module is presented to learn the feature relationship and fully utilize the extracted features by introducing two paralleled attention subnets Squeeze-and-Excitation module and Non-local module. PA-Module can help improve the performance of detecting and classification by emphasizing useful features, suppressing useless features selectively and capturing long-range dependencies in images. FreeAnchor is adopted to deal with the restriction of hand-crafted anchor assignment according to Intersection-over-Unit. It can help find the best anchor for each object by learning, and improve the performance of detecting slender objects and the ones in crowded scenes. On the public dataset OPIXray, CFPA-Net achieves 85.82% detection mean Average Precision. Moreover, achieving 81.61% classification mean Average Precision on the SIXray10 dataset. The experimental results show that our proposed CFPA-Net is more accurate and robust for the X-ray baggage inspection with densely occluded objects and complicated backgrounds.