X-ray Security Inspection Images Classification Based On Deep Learning

X-ray security inspection has always played a vital role in maintaining transportation safety and protecting public places from security threats.The rapid and accurate classification of X-ray inspection images is even more important.However,the existing X-ray security inspection work largely relies on manual operation,so the efficiency is low and the labor cost is high.Today,with the rapid development of artificial intelligence,combined with deep learning algorithms,it is of practical significance to achieve high-accuracy rapid and fault-tolerant threat objects recognition in X-ray security images.In view of the characteristics of X-ray security inspection images,such as unbalanced samples,messy background,serious overlap,and multi-scale dangerous goods,this paper studies the X-ray security inspection image classification method based on deep learning.The main research contents of this paper include:First,the design and implementation of the X-ray security inspection image classification algorithm OctConv-ABiGRU based on the combination of Octave Convolution(OctConv)and with attention-based bidirectional Gated Recurrent Unit(BiGRU)neural network.First,the method of random oversampling of threat objects is adopted to increase the representativeness of minority samples and reduce the impact of imbalanced sample on the model.Second,the octave convolution is used to replace the traditional convolution operation to divide the input feature vector into high and low frequencies,which reduces the resolution of low frequency features,reduces spatial redundancy and effectively extracting the features of X-ray security inspection images.Thirdly,the bidirectional gated circulation unit neural network with attention mechanism is used to construct different feature sequences according to different categories,adjust feature weights dynamically by learning,and improve the classification accuracy of the model for each category of thread objects.Compared with the other convolutional neural network,this method can effectively improve the accuracy and speed of X-ray security inspection images classification of threat objects.Then,the design and implementation of the X-ray security inspection image classification and location algorithm ASPP-YOLOv4 based on the combination of Atrous Spatial Pyramid Pooling(ASPP)and YOLOv4.First,the YOLOv4 network pre-trained on VOC data set was used for parameter migration,so that the model could better learn the characteristics of thread objects.Second,the ASPP module of paralleling more dilated convolution with different dilation rates was used to increase the receptive field of different sizes,so as to better capture the context information and increase the adaptability of the model.Third,the K-means clustering method is adopted to generate initial candidate boxes that are more suitable for thread objects detection.In addition,cosine annealing learning rate is adopted to further accelerate model convergence and improve model detection precision.Experiments have proved that this method can effectively reduce the omission rate of X-ray security inspection images and improve the identification ability of small target thread objects.