Pneumatic systems are extensively utilized across various industries due to their cost-effectiveness and cleanliness. Surface defects in cylinder piston rods can lead to various issues within pneumatic systems, such as reduced control accuracy and leakage. Consequently, accurate detection of surface defects in cylinder piston rods holds significant importance for the performance and reliability of pneumatic systems. However, distributional discrepancies among surface images are inevitable owing to environmental conditions. Because deep learning methods are sensitive to differences in data distribution, the performance of deep learning methods will be degraded. To tackle this challenge, an Enhancement-Segmentation-Decision Network (ESDN) is proposed for defect identification with differences in data distribution by combining image enhancement and deep-learning methods. Specifically, ESDN adopts image enhancement methods to reduce distributional differences among surface images, where image enhancement methods can improve the contrast of the defect area, and balance image grayscale or color to make defects more visible. Next, ESDN utilizes the Segmentation Decision Network (SDN) to detect and localize defects within the enhanced images. Finally, sufficient results demonstrate that the ESDN addresses the issue of distribution discrepancies in surface images, and improves the accuracy of detecting and localizing defects from 0.8955 to 0.9793.
Lu et al. (Wed,) studied this question.