| Recent years have witnessed the breakthrough of the research on deep neural networks,however,putting the potential of artificial intelligence into practice remains a great challenge.Though deep convolutional neural networks achieved the new state of image recognition on the Image Net dataset,they are still immature for commercial application.This is not only due to the complexity of understanding an image but also the massive computational budget brought by the image processing.The latter has aroused the heterogeneous computing platforms with GPU to accelerate the training of deep learning technology.Meanwhile,huge investments in copyrights and network bandwidths urge online video providers to balance the user experience and business income.To address the issue of recognizing and locate the specific object category in a video,this thesis adopt a feasible approach combining image segmentation and convolutional neural networks based on heterogeneous computing with Open CL.The solution proposed consists of three modules:(1)Video images segmentation: preprocess source videos to get frame images then apply efficient graph-based image segmentation together with selective search to locate the bounding boxes of objects;(2)Optimization of the Caffe's cross-platform character: replace the CUDA modules of Caffe restricted on NVIDIA GPU with Open CL,an open,standard framework for heterogeneous system;(3)Convolutional neural networks trained to carry out image recognition: deploy trained convolutional networks models using Image Net on Caffe,filter objected bounding boxes and trace the region back to source video.This thesis also conducted experiments to verify the feasibility and validity of the proposed solution.The results demonstrated in diagrams show that the solution is capable of locating the target objects.Besides,digging into the inner structure of convolutional neural networks helps researchers know more about mechanisms of deep learning and inspires commercial applications. |
PDF Full Text Request