Research On YOLO Object Detection Algorithm And Its Application

Object detection is one of the most important techniques in computer vision,and its main task is to locate and identify objects in an image or video.With the continuous development of deep learning,object detection algorithms based on convolutional networks have achieved better results than traditional image algorithms and are widely used in many fields.Object detection algorithms based on convolutional networks often lead to reduced detection accuracy due to occlusion,overlap,small and dense objects,and have problems such as wrong detection,missed detection and re-detection.By optimising the neural network model and its training algorithm,the detection accuracy can be effectively improved.To this end,this thesis investigates the You Only Look Once(YOLO)object detection algorithm improvement and its application,and the main work and innovations in this thesis are as follows:(1)The Deepwise Grouping Separable Convolution Detector(DGSC-Det),a YOLO object detection algorithm based on depthwise separable convolution,is proposed.The algorithm first optimizes deepwise convolution to deepwise grouping convolution in order to improve its feature extraction capability.Then,an optimised depthwise separable convolution is added to the YOLO-Head to enhance the localisation and classification capabilities of the network.Experiments were conducted on the Visual Object Classes(VOC)2007 and VOC 2012 datasets to compare this algorithm with its counterpart,and the results showed that DGSC-Det has higher detection accuracy.The algorithm is applied to the problem of detecting mask protection against respiratory diseases such as COVID-19,with good results.(2)The Differential Spatial Attention Mechanism Detector(DSAM-Det),a YOLO object detection algorithm based on the spatial attention mechanism,is proposed.The algorithm first considers the variability between feature map channels and assigns different weight matrices to different channels in order to improve the ability of the spatial attention mechanism to retain valid information and suppress noisy information.Then,an optimised spatial attention mechanism is added to the backbone network of the YOLO algorithm to enhance feature extraction.Finally,the optimised network model was subjected to computational costing statistics.Experiments were conducted on the VOC 2007 and VOC 2012 datasets to compare this algorithm with its counterpart,and the results showed that DSAM-Det has higher detection accuracy.The algorithm is applied to the problem of detecting helmet protection in cycling accidents,with good results.