Fast Deep Neural Networks With Knowledge Guided Training And Predicted Regions Of Interests For Real-time Video Object Detection

With attaching importance to deep learning methods,many applications have been developed innovatively in daily life,such as the applications of object detection.However,it suffers from extremely high computational complexity and requires expensive hardware to supporting its calculation so that it is importance that speeding up the speed of object detection while keeps accuracy of detection in practical applications.In this paper,it develops a fast-deep neural network for real-time object detection by exploring the ideas of knowledge-guided training and predicted base on region of interests.In training part,a low-complexity network that trained by the knowledge-guided training performed high accuracy,while computational complexity was reduced using the region of interests in testing part.The main works of paper are as follow:1.Developed a knowledge-guided training method based on a projection layer.Analyzed the fast-convolutional neural network and matrix multiplication which are the basic part of convolutional network.The knowledge projection layer bridges two models that guiding the training of thinner networks by large network,achieving better network performance.2.Verified the stability and accuracy of knowledge-guided training method with PASCAL and CIFAR-10 databases.The testing result shows that this method keeps the low-complexity of model,but also improves accuracy and testing time of training part.3.Proposed a framework that achieved object detection based on the region of interest.Because convolutional layer is the most complicated part of calculation,this method significantly reduces the computational complexity by filtering the redundant weights and keeps high performance.4.Constructed a fast video object detection system based on a traffic database,which extracts object features and its histogram of gradient to verify the accuracy of the detection framework.The region of interest that contain the target objects with high confidence guides the calculation of convolution.In addition,this system results on object detection from video demonstrated that the proposed method can speed up the network while maintaining the detection performance.