Detection Of Prohibited Items In X-ray Image Based On Fully Convolutional Network

The main task of object detection is to identify the classification and location information of the specific object in the image,which is also one of the core tasks in the field of computer vision.In X-ray images,objects of different materials,thicknesses and hardness are represented by pixel values,so the images are transparent and overlapped.Generally,the collected images are very crowded and disordered,and the power of X-ray machine produces interference noise and other factors,which makes the detection of prohibited items in X-ray images more challenging than other detection work.In order to improve the efficiency of civil aviation security inspection,an algorithm is proposed to quickly detect the prohibited items such as guns,ammunition,controlled knives and so on in the output image of the security detector.By this way,the security inspectors only need to open and screen the bags with suspicious items,so as to improve the detection efficiency and enhance the privacy protection of passengers.Based on this,an improved algorithm for the detection of prohibited items in X-ray images is proposed,which includes the following three parts:(1)In order to reduce the missing detection rate of prohibited items in X-ray images,a object detection algorithm based on full convolution network is proposed.According to the characteristics of X-ray image which is transparent and overlapped,the algorithm uses full convolution network as the basic network to detect the object in the way of pixel-by-pixel prediction.The structure of the network is to replace the last fully connected layer in the convolution neural network with the convolution layer for calculation,which eliminates the pre-defined anchor frame.In this way,the complex IOU calculation and matching problem in the traditional algorithm based on anchor frame is avoided,and the computational efficiency of the model is improved.(2)For the purpose of solve the problem of fuzzy border labeling caused by overlapping items in detection,a multi-level prediction is constructed to improve the recall rate.Firstly,in the feature fusion part,convolution and up sampling are performed to construct bidirectional feature pyramid,and then different boundary box regression ranges are defined in each scale feature layer.Finally,the feature graph is used to separate the objects with different sizes for attribute prediction and model training,and the detection results can be obtained.(3)Aiming at the problem of low training performance of detection model and regression of low-quality samples,an improved loss function is proposed.By introducing a weighted penalty term based on the regression loss function,the center point of the prediction frame can be closer to the object frame to a greater extent,eliminating redundancy and making the detection model converge quickly in training.The experimental results show that the average omission rate,average precision and other performance indicators of the proposed detection algorithm are improved in the self-made dataset of civil aviation security inspection,and can be better applied in the actual security inspection scene,which proves the effectiveness of the X-ray image prohibited items detection algorithm based on full convolution network.