Research On Improved Object Detection Based On YOLO

Object detection is to solve the comprehensive problem of what and where objects are in an image or a video.It is the cornerstone of more complex computer vision tasks such as object segmentation,object tracking and behavior recognition.In recent years,the continuous improvement of computer computing power has enabled deep learning technology with powerful feature expression ability to achieve significant breakthrough in computer vision,natural language processing,and other fields.The object detection algorithm based on deep learning has become the mainstream of research.YOLO(You Only Look Once)is an object detection algorithm based on deep convolutional neural network,which achieves real-time detection speed while maintaining high detection accuracy,which is suitable for engineering fields.There are four versions of YOLO since it was proposed.The principles of YOLO series are studied in this article.Improved work on Tiny-YOLOv2 and Tiny-YOLOv3 was carried out.(1)The main problem of the YOLO is the accuracy is not high,and the recall rate is lower than the object detection method based on region proposal.In YOLOv2,a set of prior bounding boxes are obtained by using K-means,improving the algorithm with higher recall rate.In this paper,K-means ++ is used to optimize the anchor box selection strategy in Tiny-YOLOv2 to improve the Avg IOU between the ground truth and the predicted box.An optimized Spatial Pyramid Pooling(SPP)module is added to Tiny-YOLOv2.Different from the traditional SPP module designed for the fully connected layer,the SPP module designed in this paper contains three max pooling layers of different sizes.The SPP module performs multi-scale transformation on a layer of the network's feature maps containing complex semantic features,and finally combines the multi-scale feature maps into a single feature map to achieve multiscale utilization of local features.Compared with the original model,the improved model improves the recall rate and detection accuracy with increasing the model complexity a little.(2)The backbone used for feature extraction in the object detection algorithm plays an role in affecting the detection accuracy and speed of the algorithm.The YOLO series have been constantly improving the backbone.The feature extraction network of Tiny-YOLOv3 is constructed by some simple convolutional layers.There is not much information exchange between the features of each level.In addition,the large number of convolution kernels in the last layer of the backbone makes the model with parameters.A multi-scale and multi-target detection method which introduces the concept of dense connection in DenseNet is proposed in this paper.The effect of the number of different sub-blocks on the experimental results was compared.A dense connection module with five sub-blocks is built to optimize the feature extraction network of Tiny-YOLOv3.The module is able to strengthen the feature reuse in the pre-propagation and post-propagation process of the network,thereby enhancing the features reuse and improve the accuracy of network model detection.On the PASCAL VOC dataset,Dense-Tiny-YOLO in this paper is superior to Tiny-YOLOv2 and Tiny-YOLOv3 in terms of model complexity,model size,detection speed,and accuracy.On the MS COCO dataset,the improved model has made great improvements in the detection of both medium and large targets.Experimental results show that Dense-Tiny-YOLO can effectively improve the performance of object detection.