# DEENet: an edge-enhanced CNN–Transformer dual-encoder model for steel surface defect detection

## TL;DR

DEENet is a new model for detecting defects on steel surfaces using enhanced edge detection and dual-branch feature extraction.

## Contribution

The novel DEENet model introduces edge enhancement and dual-encoder architecture for improved steel defect detection.

## Key findings

- DEENet achieves 81.4% mean Average Precision on the NEU-DET dataset.
- The model improves detection of small defects and enhances edge information.
- Dual-branch feature extraction and fusion improve detection accuracy.

## Abstract

Steel is an indispensable material in modern industry, and its surface quality directly affects the performance and service life of products. To address problems of insufficient feature extraction capability, weak detection of small defects, and blurred target contours that lead to degraded edge information in steel surface defect detection, this paper proposes a novel edge-enhanced dual-branch steel surface defect recognition model, DEENet. First, a dual-encoder module based on CNN and Transformer is designed to extract image features and enhance the feature extraction capacity of the backbone network. Second, a Dual Channel Fusion module is introduced to perform cross-enhancement between the local features captured by the CNN and the global semantic features modeled by the Transformer, achieving feature complementarity and improving the detection accuracy for small defects. Finally, an edge enhancement module, C2f_EEM, is designed to highlight gradient differences between defective and normal regions through differential operations, thereby strengthening contour information and improving the model’s sensitivity to defect edges. Experimental results on the NEU-DET dataset show that, compared with other algorithms, DEENet achieves a superior mean Average Precision (mAP) of 81.4%, enabling more accurate detection of steel surface defects and providing valuable reference for defect inspection in real-world production scenarios.