RDA-IC: a data synthesis approach for enhancing network performance in image classification tasks
Qi Gao, Yao Li, Yi Ping Du, Jiang Jiang, Haiyue Yu · 2025
This paper proposes a novel data synthesis method to enhance the performance of image classification tasks. The method begins by transforming raw image data into vector representations using a feature extractor, followed by clustering the features using the DBSCAN algorithm to automatically determine cluster divisions. Subsequently, the relationships between adjacent clusters are modeled as a Traveling Salesman Problem (TSP) to find the shortest path connecting all cluster centroids, optimizing the order for subsequent data interpolation. Linear fitting is performed between sample points from adjacent clusters to characterize inter-cluster feature relationships, and matching results are generated. Using a multi-stage minimum weight matching algorithm, sample pairs are determined, and linear interpolation is applied to generate new feature data and corresponding label data. A hyperparameter is introduced to control the number of interpolated samples per unit distance, enhancing the diversity and representativeness of the synthesized data. Experimental results demonstrate that the proposed method significantly improves the quality of the dataset and enhances the performance of classification models.