Comparative Study on Image Classification Models of Ancient Ceramic Types

One of the main problems faced by image big data is how to efficiently automate image classification. The current mainstream approach is to use image recognition models for training. Although there are many different image recognition models applicable to the field of image classification at present, there is a lack of comparative research on multiple neural network models for specific usage scenarios. This article focuses on the recognition and classification of ancient pottery and porcelain types, and compares and studies three mainstream image recognition models in the industry. Multiple batches of dual parameter optimization training schemes are used to obtain multiple sets of comparable training results. The test data shows that the accuracy of the three image recognition models (Inception, ResNet, MobileNet) is up to 90%, and they can effectively recognize ceramic types. Under the same hardware conditions, the MobileNet model has the fastest recognition speed, with an average of 21.35ms per image, and its computational power requirement (FLOPS) is only 0.56G, far lower than the other two recognition models, making it suitable for mobile internet application scenarios such as mobile phone image recognition.