| Object detection refers to identifying and locating specific types of targets in images or videos,and is widely employed in scopes such as face recognition,autonomous driving,intelligent security,and medical image analysis.Therefore,the research and optimization of object detection algorithms have great practical value.The intention of this article is to intensify the representational ability of object detection network and make it learn features more exhaustively,thereby improving the accuracy of algorithm.Therefore,some improved strategies are proposed.Based on widely used YOLOv3,on the network structure,the spatial pyramid pooling is embedded to strengthen the ability of feature extraction in the local area,and increasing the receptive field of feature map through integrated dilated convolution.Besides,dual attention is used to improve the quality of feature expression;In the learning strategy,the adaptive anchor box is used instead of the K-Means algorithm to make the size of bounding box more appropriate,and improving the model’s robustness against noisy background by data augmentation of mixed pixel.Additionally,weighting factor is added to balance the quantity of different categories of samples,and the loss is calculated based on intersection over union to improve the positioning effect of the model.The algorithm model is established,trained and evaluate under the Pytorch deep learning framework,and results on the MS COCO(test-dev 2017)data set show that using these improved strategies proposed in this article comprehensively makes the accuracy of YOLOv3 algorithm increase by 10.2%.Among them,new feature fusion strategy and calculation method of loss function improved accuracy of the algorithm the most,which increase the accuracy by3.6% and 2.7% respectively.Compared with other current mainstream one-stage object detection algorithms,the improved algorithm has better accuracy and speed advantages.Experiments and analysis show that these improved strategies proposed in this article can improve the effect of one-stage algorithms of object detection on medium and large objects,and increase the positioning ability of the model,which make algorithm more accurate and meet the requirements of the efficiency in life. |
