Research on the application of high-efficiency detectors into the detection of prohibited item in X-ray images

摘要

X-ray imaging can be used to inspect the internal structure of the objects without destruction, so visual inspection based on X-ray images is widely used in the security check such as customs, airports, railway stations, and postal express. Especially in the postal express industry, fast and accurate inspection of express parcels can effectively improve logistics efficiency. This article studies the application of computer vision technology to detect prohibited items in X-ray images. Due to the multi-pose objects in the packages under multi-views, it is difficult to find out the prohibited item from the packages under a single view. This article explores how to solve this problem with the loss function of classification and the attention mechanism of convolutional neural network, and apply them to high-efficiency detectors. On the one hand, we proposed a new loss function named truncated loss for X-ray image classification task. In the proposed loss, we truncated input vector of loss layer to reduce the difference within the intra-classes and increase the difference between the inter-classes. On the other hand, we proposed two new architectures for the high-efficiency detectors for the purpose of obtaining the visual features of prohibited item more effectively. One of the new architectures named channel context block (CC block), and it is based on global context (GC block). It contains global context information on each channel through operations of global average pooling, which is different from global context (GC) block. The other one of the architectures named GCC block, it is formed by merging channel context block (CC block) and global context (GC) block, and it is used to further improve the detection accuracy of prohibited item. The results of experiments on the currently widely used high-efficiency detectors in GDXray dataset show that our proposed truncated loss can improve the detection accuracy of prohibited item to a certain extent, and the new architectures can improve detection accuracy to a greater extent. The algorithms proposed in this article are also state-of-the-art on GDXray dataset.