Research And Application Of Lightweight Object Detection Based On Deep Learning

Object detection is a basic research topic in the field of computer vision.At present,it has been widely used in many fields such as transportation,security,medical treatment and so on.In practical applications,some object detection models often need to be deployed to devices such as mobile phones or embedded edge devices.However,the storage capacity and computing resources of such devices are very limited.To solve these problems,this thesis focuses on the research of the lightweight object detection algorithm based on deep learning,which not only ensures the accuracy,but also reduces the amount of model parameters and complexity.The main work and contributions are itemized as follows:1.For the YOLOv4,an improved iterative pruning method is used to prune its channel.This thesis introduces a dynamic initial penalty factor in the sparse training stage to reduce the loss of accuracy.In the pruning stage,a reasonable pruning mode is set for the special structure,and the global and local pruning thresholds are given to better control the pruning rate.The experimental results on the VOC dataset show that the pruning method can greatly compress the model size and the num of parameters to 4% of the original model.The m AP is only 1% lower than that of YOLOv4-tiny.2.A lightweight detection network named GSDet is proposed.The first 13 layers of Ghost Net are intercepted as the backbone.The optimized lightweight PAN＿Slim is adopted for feature fusion.The detection head uses depthwise convolution,and adopts anchor-free detection to reduce redundant calculation.The m AP of the GSDet on the COCO val set is21.7%,the num of the parameters is 1.19 M,and the FLOPs is only 0.51 G.The results effectively prove that GSDet still has a good detection performance under lower complexity.3.To further improve the performance of YOLOX-Nano,three optimization strategies are introduced.In the training stage,this thesis adjusts the scope of data augmentation,and the focal loss and CIo U loss are adopted respectively for classification and regression to alleviate the data imbalance and accelerate the convergence of the model training.The improved YOLOX-Nano model increases the m AP on the VOC and COCO dataset by 0.6%and 0.7% respectively,and fortunately it does not need additional inference overhead.4.Based on the ncnn framework,the improved YOLOX-Nano model is applied to the Android platform.The deployment of the detection model on the mobile terminal is realized,and the FPS on the test phone is 26,which achieves the real-time detection.