Research On Improved YOLOv3 Object Detection Algorithm And Its Lightweight Network

Object detection is an important research issue in the field of computer vision,which is mainly used to identify and locate specific objects.With the continuous development of computer hardware and deep learning,object detection algorithms based on deep learning are increasingly applied in these areas,such as intelligent security,automatic driving,and intelligent traffic processing.Therefore,this thesis further improves the YOLOv3 object detection algorithm based on deep learning to promote its detection accuracy.Then,a method of lightweight network replacement is adopted to reduce the above model's size,which makes the application scenarios of the proposed algorithm more extensive and the detection process more efficient.The specific research contents of this thesis are as follows:In this thesis,a new detection sub-network model structure is designed to solve the problem of poor detection ability of YOLOv3.It fuses the output of the hybrid dilated convolution on the second residual block in Darknet-53 and the output of the 8x downsampling of the original network.In addition,Focal Loss is used to compute the confidence of negative samples of the loss function,which can alleviate the imbalance problem of positive and negative samples of YOLOv3.Experiments on NWPUVHR-10 dataset and PASCAL VOC dataset show that the improved YOLOv3 has better detection performance compared with the YOLOv3 algorithm,and the detection speed has not dropped significantly.The improved YOLOv3 has better detection accuracy,but there are some problems such as too much model storage space and too many network parameters,which leads to limited deployment equipment.In this thesis,a lightweight method for the above model is proposed to solve this problem.First of all,the lightweight convolutional neural network EfficientNet-B0 is used to replace the backbone network Darknet-53 of the improved YOLOv3,which has achieved a significant reduction in model storage space and network parameters.Secondly,the new detection model structure of the above improved algorithm is deleted and restored to the previous detection structure of YOLOv3.Finally,the YOLOv3 detection network has continuous hybrid convolution modules with convolution kernel sizes of 1×1 and 3×3 before outputting the final predicted features.In this these,CSPBlock is used to replace the hybrid convolution module,which can further reduce the number of model parameters and simplify the model size.Experiments on PASCAL VOC dataset and RSOD dataset show that compared with the original model,the improved lightweight model has a great decrease in storage space and number of parameters,and the detection speed is also greatly improved,and the detection accuracy is only slightly decreased.