The Object Detection And Recognition Algorithm In Complex Scenes Based On Deep Learning

Nowadays,artificial intelligence is developing with an unbelievable speed.As an important branch,computer vision has become the focus of research.With the development of deep learning,computers have surpassed humans in terms of image recognition accuracy.Object detection and recognition in complex environment mainly include two parts,scene detection and object recognition.Scene detection means suspected object location,and its main methods are sliding window method and selective search method.The mainstream method of object recognition is convolution neural network(CNN),including some methods combining object location and deep learning,such as Fast R-CNN,SPP-net and so on.However,although the recognition accuracy of CNN is higher than traditional methods,training CNN models needs big data,long training time and expensive hardware.These shortcomings limit the application of CNN.In addition,as the basis of object recognition,scene detection also has many problems such as high complexity,long location time and low positioning accuracy.In order to solve the problems above,we research the object detection algorithm in complex scenes based on deep learning and its algorithm application.Firstly,we introduce the research background and progress of object detection and deep learning,including CNN and window location methods.Then we research the image saliency detection method(ISDM)based on compressed sensing(CS),using the characteristic of CS that it can collect and compress the data synchronously.In ISDM,we determine the saliency of image blocks in CS measurement domain and generate saliency block map to remove most image background redundancy.Then we propose window location method based on ISDM,which can reduce window scanning time greatly and solve problems of high image redundancy and long location time.Then we propose an object recognition method based on fast feature fusion(FFF).In this method,we design a fast feature fusion network.This network mainly includes two modules,namely feature fusion module and partial selection module.In feature fusion module,we design a shallow CNN model and then deduce the fitness function and coding rule of genetic algorithm(GA).We fuse deep learning features and traditional image features together by GA.In partial selection module,we reduce the dimension of fusionfeatures and average inner class distance,which can improve final recognition accuracy further.This method can solve model training problems of long training time and high hardware cost.At last,we achieve a deep learning web app for specific vehicle detection using the algorithm in this paper,using bayonet HD camera data.We use B/S network architecture designing this web app,which completes the application of deep learning in big image data.This app can quickly recognize specific vehicles from multipath camera data in a short time,and we can visit its query interface by browsers in phones or personal computers.Experimental results show that,the web app we developed has the characteristics of high stability,high recognition accuracy,high response speed and high scalability,which can be used in the areas of smart transportation and digital city.