An Improving Object Detection Algorithm Based On Deep Convolutional Neural Networks

Object detection is one of the most difficult and fundamental tasks in the field of image processing.In recent years,as a breakthrough in artificial intelligence community,deep learning has dramatically improved the state-of-the-art in image processing and many other domains.In this thesis,the shortcomings of the traditional object detection algorithm are described firstly,and then the advantages of the object detection algorithms based on the deep learning are analyzed.What's more,a series of object detection algorithms based on deep convolution neural networks over the past three years are summarized in the thesis.The thesis also introduces the optimization directions of deep learning network.In order to further increase the object detection speed and reduced the storage of network,this thesis improves a fast target detection framework(SSD)so that the network can be deployed on storage-limited mobile devices or embedded devices.The improvements are attached as following:1.The feature extraction network VGG16 is replaced with Squeeze network.2.The traditional convolution modules are replaced with fire module in the multi-scale feature layer.3.The bypass connection modules are added in the feature network.4.A reverse cascade rejection classifier is proposed,which is combined with multi-scale feature layer and can reduce the candidate regions to accelerate detection.5.Network pruning and weight quantization are used to reduce the model size.The thesis also carried out a series of experiments to test the impact of the above improvements.Compared with the SSD300,at the same precision level,the detection time of the improved model is about one-third of the SSD300,and the model size is reduced to above 21 times,so the new model has greatly increased the efficiency.