# Steel surface defect segmentation with SME-DeeplabV3+

## TL;DR

This paper introduces an improved deep learning model for accurately detecting defects on steel surfaces, enhancing inspection efficiency and accuracy.

## Contribution

The novel SME-DeepLabV3+ model integrates StarNet, ELA, and MSAA modules for enhanced steel defect segmentation performance.

## Key findings

- The model achieved a 1.65% increase in mIoU, 2.19% in precision, and 0.36% in MPA compared to traditional methods.
- The integration of StarNet and MSAA modules improved defect detection accuracy while reducing computational costs.
- Dynamic threshold adjustment and multiscale attention mechanisms reduced missed detections and false positives.

## Abstract

Accurate segmentation of steel surface defects is crucial for ensuring steel quality. This paper presents a steel surface defect segmentation method based on SME-DeepLabV3+ to improve the accuracy and efficiency of segmentation. First, StarNet is adopted as the backbone network, whose unique star operation can achieve efficient transformation from low-dimensional space to high-dimensional features, enhancing the model’s ability to capture steel defect features and accurately distinguish between normal and defective areas. Second, the ELA module is introduced, which is based on an efficient local attention mechanism and uses different sizes of convolution kernels for multiscale analysis of feature maps. During training, it adaptively initializes the weights of convolutional layers and introduces a dynamic threshold adjustment mechanism to adjust thresholds in real time according to the defect conditions of training batches, reducing missed detections and false positives. Finally, we integrate the MSAA module from CM-UNet, whose multiscale attention mechanism can dynamically adjust attention allocation based on defect size, avoiding detection omissions or misjudgements caused by size differences. The experimental results show that the improved model performs excellently in steel surface defect segmentation tasks, significantly improving accuracy and efficiency compared with traditional methods. The mIoU, precision, and MPA evaluation metrics increased by 1.65%, 2.19%, and 0.36%, respectively, providing more effective technical support for steel quality inspection. The combination of StarNet with the MSAA and ELA modules effectively enhances the performance of semantic segmentation models in steel defect detection while reducing computational resource requirements. The code is available at https://github.com/Eric-863/SME-main/tree/main.