Optimization of target detection algorithm based on deep learning in long tail distribution data

The long tail distribution data is characterized by a large number of samples in a few categories and a scarcity of samples in most categories, which challenges the traditional target detection algorithm, resulting in the model tending to pay too much attention to the head class with rich samples and ignore the tail class with scarce samples. This paper presents an Adaptive Class-Balanced Network (ACBN), which aims to enhance the model's recognition ability for tail classes while maintaining high performance on head classes through three key components: an Adaptive Feature Fusion Module (AFFM), Dynamic Class Weight Adjustment (DCWA), and a Class-Balanced Loss Function (CBLF). AFFM dynamically assigns the weights of different class features through attention mechanism, and enhances the extraction of tail class features; DCWA dynamically adjusts the weight of categories in the loss function according to the number of samples, giving greater attention to tail categories; CBLF combines DCWA strategy to optimize the performance of the model on long tail distribution data. The experimental results show that the performance of ACBN on LVIS and OpenImages data sets is significantly better than other mainstream target detection methods, and the indicators such as accuracy, recall, mAP and F1-score all show the superiority of ACBN in processing long tail distribution data. In addition, the ablation experiment verifies the importance of each component in ACBN, and shows that integrating these three components is very important to achieve the performance balance between categories. This study not only improves the performance of the model in complex scenes, but also provides a new optimization strategy for deep learning target detection on long tail distribution data.