Multi-Scale Information Fusion Target Recognition Algorithm for Complex Scenes

Due to the varying scale and diversity of target objects in complex scenes, multi-target recognition systems are prone to encountering issues such as false detection and low accuracy. In response to these challenges, we improve and optimize the multi-target recognition model from the perspective of multi-scale information fusion and propose RBs-YOLOv8. Based on the YOLOv8 framework, this model combines the RepViT framework for feature extraction and optimizes the cross-scale connection in feature fusion. At the same time, the BiFPN module is used for weighted fusion. For the fusion of varying channel information, we also propose an attention mechanism to enhance the model's capacity for learning global features. It embeds it into the splicing process of features of various scales. Through training and verification on data sets in real scenes, experimental results show that compared with the traditional YOLOv8 model, the average accuracy of the improved RBs-YOLOv8 model is improved by 17 percentage points, and the loss function decreases faster and smoother during training. Additionally, To determine how each improvement contributes to the model's ability to learn global features, we conduct a series of ablation experiments. Demonstrating the efficacy and proving its effectiveness, the experiment results showcase the proposed multi-target recognition algorithm.