Improved image semantic segmentation with domain adaptation for mechanical parts

Abstract

Non-contact detection methods based on computer vision are widely used in industrial production. When collecting images, different light source schemes are often used to meet the detection requirements of different objects. The image styles collected under different lighting scenes are diverse, and the image semantic segmentation model trained with a specific dataset has unsatisfactory performance when processing images in different domain. This paper designs a mechanical part image semantic segmentation model based on domain adaptation and GAN. In order to verify the effectiveness of the proposed method, this paper collects and labels some gear images and constructs a dataset. The first step is to train encode-decoder structure with an enhanced memory module with source dataset to achieve semantic segmentation, then the dataset is transformed into a target domain by GAN, and finally the image semantic segmentation model is fine-tuned with the target dataset. The advantage is that since the cycle-consistency loss is used to constrain the spatial structure of the reconstructed image, the two datasets can share a semantic segmentation label. Experiments show that the image semantic segmentation method based on domain adaptation has achieved good results on different styles of gear parts image datasets, and achieved a pixel accuracy of 94.1%.

Keywords

• domain adaptation
• memory module
• encode-decoder network
• semantic segmentation
• generative adversarial networks