Multi Scale X-Ray Image Detection Based On Deep Learning

At present,intelligent security inspection is developed and studied by more and more people.But it also faces huge challenges when applied to actual scenarios.For example,the scale of prohibited items in X-ray images is very varied widely.And various prohibited items are obscured.These circumstances can lead to false detection and missed detection.And it also brings huge hidden dangers to the security inspection.Aiming at the problems of multi-scale and occlusion of prohibited items,a new generation of improved algorithm based on deep learning is proposed.(1)Aiming at the scale difference problem of prohibited items in X-ray images security inspection,a multi channel region proposal network(MCRPN)is designed on the basis of Faster RCNN.Firstly,due to the complementarity of different convolutional layers features in visual semantic,multi-level feature extraction is introduced to combine richer high-level semantic features and low-level detail information closer to the original image of VGG16.Secondly,the anchor box parameters in multi-channel RPN are modified.And objects of different scales pass through independent RPN networks.The generated multi-scale candidate regions are mapped to the corresponding feature maps.So a multi-scale prohibited items detection network is constructed.Finally,a multi-branched dilated convolution module(DCM)is designed on the multi channels to increase the receptive field.So that the feature of different scales are enhanced.The improved multi-scale network is tested on the self-made data set SIXray＿OD.The average detection accuracy reaches 84.69%,and the detection performance is 6.28% higher than the original network.The experimental results show that the recognition accuracy of the improved algorithm is improved.And the multi-scale problem is improved to some extent.(2)Aiming at the occlusion problem of objects in X-ray images,Soft Non maximum suppression algorithm(Soft-NMS)and Regression loss function are introduced on the basis of improved multi-scale network.Firstly,the Non-Maximum Suppression algorithm(NMS)in the Faster RCNN network is directly replaced with Soft-NMS.This reduces the score of the same type of object in the adjacent boxes rather than reducing the score directly to 0.Therefore,the missed detection caused by deleting adjacent detection boxes by mistake is reduced.So same kind of prohibited items occlusion problem is improved.Secondly,the location of the second stage of regression loss function is replaced with Repulsion loss function.It minimizes the loss value and makes the model achieve the best effect.The improved network is tested on the data set SIXray＿OD.The average detection accuracy reached 86.37%,which is 1.68% higher than the improved multi-scale network performance.Experimental results also show that the improved algorithm improves detection accuracy and improves the occlusion problem.The final improved network algorithm is 7.96% higher than the original Faster RCNN network detection accuracy.So experiments show that the improved algorithm has obvious advantages in improving multi-scale and occlusion problems.