Study Of Object Detection Algorithm Based On YOLOv5

Object detection is one of the most active research topics in the field of computer vision,which aims to classify and locate objects in images.Recently,due to the rapid development of deep learning,object detection has made great progress in video surveillance,autonomous driving,and remote sensing detection.Therefore,it is of significance for a comprehensive understanding of the object detection.The object detection algorithm based on deep learning employs convolutional neural networks to extract the semantic and geometric information of image,and then a feature map can be obtained to judge the category and confidence of the object on the feature map and mark the bounding box of object.In this procession,the advanced object detection algorithm can achieve a real-time detection.However,it is still necessary to improve the detection accuracy for small object detection,complex background object detection,camouflage object detection,and oversized object detection.Spired by this,this thesis investigates the object detection algorithm based on YOLOv5,and implements three improvements.Firstly,the 3-scale feature detection is modified into the 5-scale,which makes the detector can use more fine-grained features.Meanwhile,a new feature fusion structure is added,so that the feature information extracted by the improved structure is richer and more comprehensive.Secondly,the attention mechanism is added into different positions of the convolutional neural network,while the compressed excitation module in the channel attention mechanism is changed to an efficient channel attention module.Then,the effective features can be enhanced and the useless features can be weakened in the channel domain and the spatial domain during the feature extraction.Last but not least,this thesis employs the several Gaussian filters in image processing the original image and the shallow feature map in the PAN,which is conducive to the detection of super large objects.This thesis theoretically analyzes the validity of the improved scheme,and selects the object detection public dataset MS COCO and the camouflage object detection dataset COD10 K for experimental verification.Experimental results show that the improved multi-scale feature detection and feature fusion structure expands the detection range and increases the recall under the specific mean average precision.After the addition of the attention mechanism,the classification and the locating of camouflage objects and small objects is more accurate.For the large objects in the MS COCO,the mean average precision and recall rates have been greatly improved with 6.3% and 5.4% respectively.The use of Gaussian smoothing for the original image and feature map makes it conducive to the detection of super large objects.Tested on the MS COCO dataset,the mean average precision and recall increased by 2.1% and 2.7% respectively.Therefore,it is proved that the improved object detection algorithm based on YOLOv5 is feasible and achieves better detection performance.