Small Target Detection In Remote Sensing Images Based On Adaptive Multi-feature Fusion

In recent years,aerial remote sensing imaging technology has advanced rapidly,and people can obtain high-resolution optical remote sensing images through remote sensing equipment.With the gradual increase in resolution,remote sensing images can reveal more detailed features of features and have rich spatial information.Aerial remote sensing technology is widely used in military reconnaissance,urban planning,fisheries management,environmental monitoring and other fields.As the aerial remote sensing equipment is imaged from a high altitude overhead angle,the small targets in the whole image are widely distributed,the number is large,the proportion of pixels is small,the effective feature information is scarce,and the features are difficult to identify,which makes the remote sensing image small target detection algorithm urgent to make a new breakthrough.This paper improves on the YOLOv3 and YOLOv5 s algorithms to enhance the performance of small target detection in remote sensing images.The main research contents of this paper are as follows.(1)To address the problem that it is difficult to detect small targets in remote sensing images,an improved YOLOv3 remote sensing image small target detection algorithm,namely YOLOv3-S,is proposed.Firstly,a parallel feature extraction auxiliary network is designed to achieve the fusion of shallow features and deep features to enrich the semantic and spatial information in the feature layer,secondly,the adaptive feature selection module is used to filter the fused features to suppress the influence of interference noise on small target detection,and the cross-layer feature fusion module is used to fuse different feature layers to enhance the connection between target context information,Finally,the effectiveness of this paper’s algorithm was verified on the Pascal voc2007 and DIOR datasets.The experimental results show that the algorithm in this paper outperforms other mainstream target detection algorithms,with m AP improving by 3.6% and 2.6% on the baseline algorithm to 88.3% and 73.8% respectively.(2)In order to further balance the detection accuracy and speed of the network model and improve the efficiency of small target detection,this paper combines an improved and innovative scheme with the advanced target detection algorithm YOLOv5 s,improves the feature extraction auxiliary network through the lightweight Ghost Net model,introduces the SPP module to expand the perceptual field,adds the CIOU loss function to make the prediction frame coordinate information regression more accurate,and uses CCNet,a pixellevel attention mechanism,to enhance the connection between different pixel contexts.After extensive comparison experiments,the algorithm achieves 59.16% and 95.07% m AP on DIOR and RSOD datasets respectively,which is 6.44% and 3.35% higher than the benchmark algorithm,and the number of parameters only increases by 2.25 M,achieving the balance between detection accuracy and speed efficiently.