| At present,socio-economic development is in a state of vigorous development,people in various countries around the world are crowded,and security inspections in public places are obviously very important.With the gradual development of millimeter wave imaging technology in the field of security inspection,research on more accurate and faster millimeter wave image target detection and recognition algorithms has very important practical application value.Due to the low resolution of millimeter wave images,it is difficult to extract features manually,and the target detection and recognition method based on deep learning has the advantage of automatically extracting target features.Therefore,millimeter wave image target detection and recognition algorithm based on deep learning has become a current research hotspot.This subject comes from a scientific research project.Based on the application of millimeter wave security inspection,the millimeter wave image data set of dangerous goods carried by the human body is collected and organized.Aiming at the problem that the resolution of the millimeter wave image is low,the target with a small size is difficult to be detected,and the detection speed is slow.In this thesis,a deep learning-based millimeter wave image target detection and recognition algorithm is studied.The main work is as follows:1.For the problem that the millimeter wave image has a small data set due to the few acquisition methods,this thesis introduces the transfer learning method to realize the millimeter wave image detection.At the same time,according to the characteristics of the millimeter wave image,the area-based target detection and recognition algorithm is studied,and the detection performance of three typical area-based target detection and recognition algorithms is compared.The experimental results show that the faster region-based convolutional neural network?Faster R-CNN?has the best performance,with a detection accuracy of 79.15% and a speed of 3.9 frame?s-1.2.For the problem that the millimeter wave image has low resolution,which makes the target in the millimeter wave image difficult to be accurately identified.This thesis studies a pixel based on region of interest?ROI?based on Faster R-CNN Corresponding Faster R-CNN algorithm.Specifically,the method corresponding to the ROI pixel retains the floating-point pixel value of the ROI coordinate,and uses the bilinear interpolation algorithm to obtain the floating-point pixel value of the sampling point,which solves the problem of ROI coordinate position error caused by the rounding operation of the ROI pooling layer.The experimental results show that the detection accuracy of the improved algorithm has increased to 87.92%.3.For the problem of low detection accuracy of small targets such as folding knives in millimeter wave images,this thesis studies a new deep convolutional neural network based on Faster R-CNN corresponding to ROI pixels.Specifically,using a deep residual network?Res Net?as a feature extraction network based on Faster R-CNN corresponding to ROI pixels,and constructing an improved method of multi-layer fusion features.Thus,a deep convolutional neural network for millimeter wave images is constructed.The experimental results show that the detection accuracy of the improved algorithm is improved to 93.91%,the detection speed is 3.7 frame?s-1,and the detection accuracy of the proposed algorithm for small targets such as folding knives is increased by about 11%.4.For the problem of slow detection speed of millimeter wave images,this thesis designs an optimized network to filter target region proposal based on the deep convolutional neural network constructed above.Specifically,by designing a binary classification network,the problem of redundant quantity of target region proposal is solved,and the detection speed of the algorithm is improved.The experimental results show that the detection accuracy of the improved algorithm is 93.11%,and the detection speed is increased to about 5 frame?s-1. |
PDF Full Text Request