Design Of Light-weighted Deep Neural Networks For Object Detection

With the rapid development of deep learning and computer vision,various object detection algorithms based on Convolutional Neural Networks(CNNs)keep refreshing the performance records on different benchmark datasets.The current mainstream object detection algorithms based on CNNs include two-stage detectors(RCNN,Fast RCNN,Faster RCNN,Mask RCNN,Trident Net)and onestage detectors(YOLO,SSD,Corner Net,Extreme Net).Although new algorithms keep refreshing their performance,the parameter amount of its model is often tens of megabytes,so it is difficult to deploy such a large model in edge devices such as mobile terminals with limited computing resources.Therefore,it has become a research hot for improving the lightweightness of the detection network under the premise of ensuring a certain accuracy rate.At present,the lightweight methods of CNNs are mainly divided into three categories: design of lightweight network architecture,model pruning,and parameter quantization.This thesis focuses on the lightweight of object detection network from four aspects: design of lightweight network architecture,parameter quantization,layer fusion and pruning.The main contributions are as follows:(1)An improved SSD lightweight object detection algorithm is proposed.The algorithm redesigned the SSD object detection framework based on Res Net50 and Mobile Net and trained a model with full precision weights.Then,based on the full-precision weight model,a block-by-block quantization strategy is adopted to reduce the weight accuracy of the convolutional layer in the feature extraction layer to three values.Experimental results show that the proposed joint scheme can achieve72.54% m AP on the Pascal VOC2007 dataset.Compared with other industry-leading lightweight object detection methods,its detection accuracy is higher and the proposed model has a smaller memory space.(2)A lightweight object detection algorithm based on OSNet is proposed.In view of the large amount of SSD parameters and the need for customized hardware,this algorithm proposes a lightweight object detection network based on OSNet.Experimental results show that the proposed scheme can achieve 75.37% m AP on the Pascal VOC2007 dataset.Compared with other industryleading lightweight object detection methods,the detection accuracy is higher and the memory space occupied by the model is smaller.At the same time,the limitation of the need to customize the hardware after the algorithm quantization is improved.The algorithm may found its applications in various mobile terminals and other edge devices.(3)Two model compression methods are proposed and applied to financial statement recognition.First,a method of channel pruning is proposed for Yolo V3 algorithm.This method improves the m AP by 3%,reduces the memory by 83%,and speeds up the reasoning time by about 1.9 times.Futhermore,a layer fusion method is proposed based on the PSENet,with the compressed model,all indicators of the model are almost unchanged,the model size is reduced from 218 MB to 109 MB,and the compression rate reaches 50%.