Semantic segmentation of end mill wear area based on transfer learning with small dataset

In the milling process, the wear area of the tool is often segmented using the traditional image processing method to quantify the tool wear value. However, these methods have the disadvantages of having weak anti-noise capabilities and low segmentation accuracy. Although the semantic segmentation network can achieve excellent segmentation accuracy, obtaining enough end mill wear images to support the semantic segmentation network's training is challenging due to the high acquisition cost of wear images. As a result, this paper suggests a small sample end mill wear area segmentation method based on transfer learning and generative adversarial networks to address the issue of insufficient samples of end mill wear images. In this paper, WGAN is used to generate wear images to expand the dataset with a few samples, and the transfer learning method is used to improve the generalization ability of the segmentation network and finally achieve small sample training. This approach increases mPA by 4.46% and mIOU by 8.97% when compared to merely using the semantic segmentation network for small sample training. According to experimental findings, this method not only has high stability and segmentation accuracy but also solves the problem of insufficient end mill wear image samples. The method proposed in this paper can be effectively applied to the intelligent detection of the tool wear state, improving the accuracy and stability of the measurement of the tool wear value.