Rsearch On Object Detection Algorithm Based On YOLOv4 Remote Sensing Image

Over the past few years,object detection has become the prevailing technique in remote sensing image processing and computer vision for acquiring certain objects.This offers remote sensing images a significant role in many scenarios such as intelligent transportation,smart city,public safety,as well as military campaign.However,the relatively far and small targets in the images are suffering from few features remained in the detection layer and low detection accuracy,due to most popular detection models nowadays deepening network layers in pursuit of deeper semantic information.Moreover,the anchor frame detection technique that is utilized by most contemporary object detection algorithms can easily miss objects with large aspect ratios and lead to poor detecting performance.Firstly,to tackle the aforementioned issue regarding the low detection accuracy of small targets,this thesis developed a small-object-friendly algorithm based on YOLOv4.To start with,the algorithm fuses shallow features with deep features,to enrich the feature information of small and medium targets in the valid feature layer.Next,the algorithm added a batch normalization layer onto the original spatial pyramid pooling structure in YOLOv4,in efforts to normalize the extracted features from the backbone network and better prepare the detection model for learning target features.Experiments with the improved algorithm on the HRRSD dataset yielded 92.92% m AP for the 13 categories,which outperformed YOLOv4 by 7%.For the Vehicle category where small targets abound,the average detection accuracy was also 6%higher than that from YOLOv4.Other categories also witnessed enhanced accuracy of various extents.Secondly,to deal with the high missing rate on objects with large aspect ratios,this thesis devised a YOLOv4-based algorithm in favor of these stout targets.To begin with,the algorithm utilized target center to conduct regression of the predicted frame,which got rid of the anchor frame associated with the detection head compared to YOLOv4,and was demonstrated effective in reducing missing rate on objects with large aspect ratios.Afterwards,this thesis also proposed introducing an attention mechanism module to increase the weights of relevant channels in the valid feature layer.This will enable the model to learn more effective feature information.Similarly,experiments with the improved algorithm on the HRRSD dataset yielded 91.11% mAP for the 13 categories,outperforming YOLOv4 by 5.19%.For the Parking lot category where objects with large aspect ratios abound,the average detection accuracy was also 8% higher than that from YOLOv4.Lastly,based on the fusion of shallow & deep feature layers and modification of the spatial pyramid pooling structure,the optimal remote sensing image detection model was obtained with the addition of the attention mechanism module.Again with the 13 categories of the HRRSD dataset,an average detection accuracy of 93.0% was achieved,outperforming YOLOv4 by 7.08%.Overall,this thesis brought up a more effective model for remote sensing image detection.