Optimizing Matrix Capsule Networks for Contraband detection Research

Abstract

An optimized matrix capsule network is proposed for the problem of contraband in parcels with different poses, different sizes, random occlusion and sample imbalance. The network improves the recognition accuracy with the help of the matrix capsule network's recognition ability for object poses, and is mainly composed of a multi-branch feature extraction network and a side branch matrix capsule network, which is used to extract large and small targets; the side branch matrix capsule network uses a larger capsule convolution kernel, which is capable of detecting larger targets, and uses the operation of randomly discarding the capsules in the side branch, which makes the parameter amount to be reduced while enhancing the learning ability of the network. The region of interest of the network is obtained by using heat map approach with the help of weight back propagation mapping back to the original map to localize the contraband. Through a large number of experiments on the SIXray dataset, it is proved that the network in this paper improves the detection accuracy by 9.43% and the processing speed of the model by about 1/3 compared with the original capsule network.