Automatic Classification of Blue and White Porcelain Sherds Based on Data Augmentation and Feature Fusion

Abstract

Many blue and white porcelain are unearthed in Jingdezhen every year. The patterns on the sherds have important research significance. At present, the classification of porcelain shards is mainly based on manual work, which has the disadvantages of large workload. The use of automatic classification methods also faces complex patterns and sample sizes. In order to solve these problems, this paper proposes a new automatic recognition method based on deep learning, including data preprocessing method combined with color segmentation algorithm, a new data augmentation method FCutMix for regions of interest, a new integration strategy and the redesigned deep network model FFCNet that integrates multiple features. After experiments, the data preprocessing method, feature fusion method and integration strategy proposed in the paper can effectively improve the performance of the model by removing redundant information and adding effective features. The FCutMix method can also obtain more accurate mixed samples than the traditional CutMix. The method proposed in this paper improves the accuracy of tasks in 14 categories from 71.7% to 83.2% in a dataset containing only 373 images of porcelain sherds. In the future, this research will further design the network structure and multi-level feature fusion.