Study On Fine-grained Recognition In Remote Sensing Image Target Detection

The utilization of spatial techniques for the detection of targets in remote sensing images constitutes a pivotal topic of computer vision research.The precision of remote sensing image detection,particularly aircraft detection,is a critical factor in the acquisition of military intelligence and air transport scheduling,among other domains.Recent studies on high-resolution remote sensing image detection based on deep learning have shown limited progress in fine-grained detection of aircraft targets,with few published reports and research results.This challenge is compounded by the lack of publicly available datasets of remote sensing images.Additionally,the small size of aircraft detection targets and their high similarity in appearance make them difficult to distinguish in large remote sensing images,resulting in a high rate of false detections.In this study,we created a fine-grained dataset of high-resolution remotely sensed aircraft images MTAR.We proposed a novel detection algorithm,namely YOLOv5s-BCT,which employs a joint attention mechanism and Bi FPN based on the YOLOv5 s algorithm.The experimental results demonstrate that our proposed algorithm outperforms other mainstream detection algorithms on the MTAR dataset in terms of detection accuracy.First,we present a novel dataset for fine-grained identification of aircraft in high-resolution remote sensing images.The MTAR dataset includes 29 distinct types of military and civilian aircraft.Second,benchmark experiments and feasibility analyses are conducted on the MTAR dataset.The performance of various detection algorithms on the MTAR dataset is also thoroughly evaluated and analyzed.Third,we proposed an improved algorithm YOLOv5-BCT based on YOLOv5 for fine-grained identification of aircraft in high-resolution remote sensing images.This algorithm introduces an attention mechanism in the backbone network to enhance the focus on typical features.In the neck network,YOLOv5-BCT uses the Bi FPN structure instead of the original PANet and adds a top-level enhancement module.A small target detection head is added in the head network,and the K-means++ algorithm is adopted for anchor frame clustering.Finally,the Triple-Loss loss function is used as the classification loss algorithm.Experimental results demonstrate that the YOLOv5-BCT algorithm achieves m AP value of 91.6% on the MTAR dataset,which is a 1.7%improvement over YOLOv5.Therefore,YOLOv5-BCT can effectively enhance the fine-grained recognition of remote sensing aircraft targets.