Object Detection And Image Classification Based On Deep Learning

Object detection and image classification is not only the basis of the computer vision,but also the core of computer vision.Object detection and image classification is closely related to people's life.In recent years,due to the brilliant results of the deep learning method achieved in the ImageNet ILSVRC,more and more people get involved in the research of object detection and image classification.Big data provide unprecedented opportunities for the development of artificial intelligence.Breakthrough made by deep learning method in the field of object detection and image classification is not a coincidence.R-CNN is the first proposed method which used deep learning algorithm in the process of object detection.Firstly,the candidate regions are selected by selective search.The feature of the region is extracted from the candidate region by using depth convolution network.Then the region is divided into object and background by using linear classifier based on support vector machine.In this paper,we improve R-CNN model to achieve an object detection and image classification system based on deep learning.First,we improve the detection module by using Edge Boxes algorithm rather than selective search.Secondly,we improve the R-CNN by modifing the R-CNN network structure,which is different from the traditional hierarchical training method.Through the end-to-end training method,we improve the object detection average accuracy rate(mAP)in the VOC PASCAL data set.In addition,the obeject detection and image classification algorithm based on our improved R-CNN algorithm can reduce the buffer space in the training phase,thus save the disk space.Finally,our object detection and classification algorithm obtained 56.8 of the mAP in the PASCAL VOC data set,compared to the V5 DPM model to enhance the 10%,compared to the R-CNN upgrade 70%.In addition,previous research has mainly focused on the improvement of the detection performance,but not focused on data.However,it is also necessary for the visualization of the neural network based on convolution neural network.Therefore,in this paper,we do a lot of work on the CNN feature extraction and visualization.It can be found that,as level of the network increases,the feature acquired is more and more,the more can be summarized from the semantic features of the image.