ACS-DeepLabv3+: A Deep Learning Network for Defect Detection in Automotive Manufacturing

Abstract

Abstracts

Automotive underbody sealant defect detection is a critical task for quality control. However, conventional semantic segmentation models often struggle with challenges such as low accuracy on small defects, imprecise edge segmentation, and high computational costs. To address these issues, this paper proposes ACS-DeepLabv3+, a lightweight and accurate network. The model enhances efficiency by replacing the original backbone with MobileNetV2 and introduces a novel, empirically validated Car-ASPP module. This module improves multi-scale feature extraction by integrating depthwise separable convolutions with dual attention mechanisms: the Convolutional Block Attention Module (CBAM) and the Selective Kernel Network (SKNet). A transfer learning strategy with early stopping is also employed to optimize the training process. On our industrial dataset, which features a variety of challenging defect types, ACS-DeepLabv3 + achieves a mean Intersection over Union (mIoU) of 85.99% and a mean Pixel Accuracy (mPA) of 91.55%. With only 6.85 M parameters and an inference speed of 33.52 FPS, our model significantly outperforms the original DeepLabv3 + and other mainstream networks in both segmentation accuracy and computational efficiency, offering a robust solution for real-time industrial applications.