Research On Automatic Detection Algorithm Of Controlled Items Based On Deep Learning

The purpose of this paper is to study the automatic detection of controlled items in security inspections,thereby improving the efficiency of security inspections and reducing the waste of human resources to a certain extent.Automatic detection of controlled items can be reduced to a object detection problem.Nowadays,deep learning has achieved good results in various fields,it's application has led to rapid development in the research of object detection algorithms.This paper specifically builds the SDCI2018 dataset for controlled items.This paper focus on the application of deep learning-based object detection algorithms such as SSD and YOLOv3 in the detection of controlled items,and a series of improvements according to actual needs to improve the detection effect on special targets,security X-ray images,and finally build an automatic detection system for controlled items to meet practical applications.In the study of SSD algorithm,in order to solve the problems of missing small objects,false detection and other problems in security inspection image detection,this paper proposed a feature fusion object detection algorithm with image sub-region detection to effectively improve the small object detection effect.This paper use strategies such as angle rotation to help augment the dataset.Among augmentation strategies,since the convolutional neural network has no rotation invariance,the angle rotation process is taken offline,so that the entire dataset is multiplied,and others are online preprocessed randomly including mirror,deformation,brightness,contrast,blending,noise addition,color space transformation,and other processing methods so as to increase the diversity of data during training.Then based on the Single Shot MultiBox Detector(SSD)algorithm,using a multiscale feature fusion method,which fuses deeper feature maps and shallow feature maps,increasing the receptive field of shallow feature maps to improves the detection accuracy of small objects and using deep feature maps to detect larger objects.Finally,when the input image is larger,for example,larger than 1024 × 1024,performing sub-region detection on the objects in the picture is helpful.In the YOLOv3 study,referring to its processing method on the dataset,performing a cluster analysis on SDCI2018 to guide the setting of a more scientific anchor to facilitate convergence during training.At the same time,this paper studied the detection effect by using different loss functions for border regression.Since any neural network used for object detection can be roughly divided into two parts,backbone and head,this paper attempts to improve the detection speed by using a lightweight network structure as the backbone,and adding an effective feature enhancement module to the head to improve the detection accuracy.In order to verify the performance of the algorithm,this paper chooses two datasets to test the algorithm: VOC0712 dataset,SDCI2018 dataset.Among them,SDCI2018 is a special data set made specifically for the inspection and collection of controlled items in this paper.It includes 17,322 pictures in 7 categories,which can be used as a baseline for the research of automatic detection algorithms for controlled items.The detection accuracy of this algorithm on the VOC0712 universal dataset is 80.3%,which is 1.4% higher than the SSD algorithm;the detection accuracy on the SDCI2018 dataset is 97.9%,which is 2.2% higher than the SSD algorithm.These can accurately detect controlled items in the security picture in real time,especially for small objects in large pictures.The GIOU as a border regression loss function in YOLO can improve the accuracy by about 1%,and the RFB module can increase the mAP of YOLOv3 by 4%.