An improved surface defect detection method based on DeepLabv3+ for piston pins

Abstract To address the pressing challenges of slow computation and suboptimal edge feature extraction in current piston pin surface defect detection techniques, this study introduces a refined methodology. By leveraging MobileNetV2, a streamlined convolutional backbone network, our approach accelerates processing speeds. Concurrently, the integration of an attention mechanism ensures that the model devotes increased focus to both edge irregularities and minute defect details. A tailored loss function mitigates the prevalent issue of sample imbalance, further enhancing model performance. Empirical evidence drawn from the piston pin dataset evidences the efficacy of our improvements: an MIoU of 85.41, MPA of 88.45, and a frame rate of 32.64 FPS, marking substantial gains in both accuracy and speed relative to pre-upgraded benchmarks. This optimized method not only delivers precise and timely defect delineation with a reduced parameter count, but it also contributes a novel and efficacious strategy for enhancing surface defect detection outcomes.