Target Detection System Based On Faster RCNN

In the direction of image recognition,target detection is one of the most important research contents.In recent years,image processing algorithms based on deep learning have been widely used in different Fields such as medical treatment,smart home and transportation.The image processing algorithms based on deep learning compared with the traditional one can extract deeper character features and learn the features according to complex environment to improve robustness and accuracy of recognition.In this dissertation,the research of image recognition algorithm based on depth learning is carried out for UAV to recognize and locate different targets.Furthermore,A model of target detection network based on Faster R-CNN(Faster Regions With CNN)is built.The model can realize feature extraction and candidate region generation of targets in a deep network framework,and improve the running time of the algorithm.Firstly,this paper uses characters as the target to make data sets,and through data preprocessing,captures the back and silhouette images of 5 characters,marking the images,adding data sets through data enhancement,batch normalization and other measures to avoid over-fitting phenomenon;Secondly,feature extraction and batch generation of candidate frames are carried out on data pictures,target features are extracted by using an improved shared convolution layer for each frame of images,candidate frames are generated in batch by using a Region Proposal Network(RPN),and using a candidate region screening algorithm that based on hierarchical clustering to screen target characters in the candidate frames in advance,so that the calculation amount and time of the model are reduced,and we can further improve effect of the image recognition.Then,ROI pooling(Region of interest pooling)is carried out on candidate feature pictures with different sizes to make the size of the candidate region pictures consistent,which is convenient for subsquent network processing.Finally,identifying and locating the characters in the picture,and the candidate feature pictures are classified by the Softmax classifier to obtain the information of the target,using regression device to cope with the candidate frame feature pictures,using the non-maximum suppression algorithm to optimize the candidate frame to obtain the coordinate data of the center point of the candidate frame,which is transmitted to the UAV control system through WIFI communication module,and the final coordinates of the target are obtained through coordinate conversion to obtain the positions of the characters,thus realizing the detection of the target by the UAV.In this paper,experiments are carried out on Faster R-CNN network model before and after improvement.During the experiment,the accuracy of target detection is about 96%,which is 6% higher than that before improvement.And the time of the model training is 53 min,which is 22 minutes higher than that before improvement.The experimental results show that the recognition effect of the algorithm in this passage is relatively better.