An Improved YOLOv5 Model Based on Feature Fusion and Attention Mechanism for Multiscale Satellite Recognition
Naijun Shen, Rui Xv, Yang Gao, Chen Qian, Qingwei Chen
Abstract
Space target recognition is a crucial task in space situational awareness and a significant research area in space exploration. Recognizing space targets poses challenges due to their varying scales and reflective surface materials, resulting in a lack of reliable visual features. This paper proposes a multi-scale spatial target recognition method based on YOLOv5 to address these limitations. The proposed method overcomes the challenges by optimizing the initial anchor frame using an improved K-means++ algorithm, enhancing network stability. Selective kernel networks are introduced to mitigate complex background interference and enhance effective feature representation, even when optical features are missing. Additionally, an improved Bi-directional Feature Pyramid Network structure extracts features from small targets in deep space by fusing multi-scale features through partial cross-stage connectivity. Combining these three improvements enables high-accuracy recognition of multi-scale targets. To diversify the dataset, Unity 3D is used to simulate multi-scale targets with reflective optical characteristics, complementing existing darkroom target images. This hybrid dataset with multi-scale targets is utilized for training and testing. Experimental results demonstrate the model’s superiority in extracting critical target features. [email protected] improved by 2.55% and [email protected]:.95 improved by 4.3%. Ablation analysis further validates the effectiveness of each improved module. Overall, the proposed model substantially enhances the recognition accuracy of multi-scale targets.