Image small target detection in complex traffic scenes based on Yolov8 multiscale feature fusion

Abstract

Addressing the challenging issues in small target detection within complex traffic scenes, such as scale variation, complex background noise, and the problems of missed and false detections, this paper introduces a Multi-Scale Feature Fusion YOLOv8 (MSFF-YOLOv8) approach. Initially, building upon the YOLOv8 detection framework, an attention mechanism module is integrated for novel adaptive feature assimilation and redistribution. This innovation facilitates the effective amalgamation of multi-scale features, thereby bolstering the model's proficiency in identifying small targets and enhancing the richness of the contextual information within the output features. Furthermore, the incorporation of deformable convolution amplifies the algorithm's capacity to maintain target consistency amidst complexity. Additionally, employing a feature distillation technique permits the student model to absorb crucial feature representations from the teacher model, circumventing the detrimental effects of semantic disparities across stages. This significantly elevates the model's generalizability and robustness. Experimental validations corroborate the efficacy and superiority of the proposed method. Enhanced detection performance is achieved, effectively mitigating the challenges of small target detection in complex scenarios, such as under poor lighting conditions in traffic environments, and elevating both the accuracy and efficiency of detection.