Ceramic Microscope Image Classification Based on Multi-Scale Fusion Bottleneck Structure and Chunking Attention Mechanism

In recent years, the status of ceramics in fields such as art, culture, and historical research has been continuously improving. However, the increase in malicious counterfeiting and forgery of ceramics has disrupted the normal order of the ceramic market and brought challenges to the identification of authenticity. Due to the intricate and interfered nature of the microscopic characteristics of ceramics, traditional identification methods have been suffering from issues of low accuracy and efficiency. To address these issues, there is a proposal for a multi-scale fusion bottleneck structure and a chunking attention module to improve the neural network model of Resnet50 and perform ceramic microscopic image classification and recognition. Firstly, the original bottleneck structure has been replaced with a multi-scale fusion bottleneck structure, which can establish a feature pyramid and establish associations between different feature layers, effectively focusing on features at different scales. Then, chunking attention modules are added to both the shallow and deep networks, respectively, to establish remote dependencies in low-level detail features and high-level semantic features, to reduce the impact of convolutional receptive field restrictions. The experimental results show that, in terms of classification accuracy and other indicators, this model surpasses the mainstream neural network models with a better classification accuracy of 3.98% compared to the benchmark model Resnet50, achieving 98.74%. Meanwhile, in comparison with non-convolutional network models, it has been found that convolutional models are more suitable for the recognition of ceramic microscopic features.