Steel plates used in aerospace, automotive, and civil infrastructure require strict surface quality control because minor defects such as crazing, inclusions, and scratches can undermine mechanical integrity and long-term safety. Traditional inspection approaches—manual visual checks and nondestructive testing (eddy current, infrared thermography, magnetic flux leakage)—have limitations in real-time performance, robustness to noise, and generalization to complex surface irregularities. The authors position deep learning–based vision systems as the contemporary solution but note three persistent gaps in prior work: sensitivity to fixed illumination and limited adaptability, generative models that experience mode collapse or insufficient fidelity particularly for small/low-contrast defects, and lightweight detectors that sacrifice fine-grained feature capture needed for small-defect detection.
The paper aims to address these gaps by designing an integrated pipeline that pairs adaptive image enhancement, attention-augmented generative sample synthesis, and a lightweight yet feature-preserving detection model tailored to industrial deployment constraints.
The proposed framework comprises three coordinated modules aligned into an end-to-end detection pipeline. The modules are:
The study evaluates the contributions of each module via ablation studies and compares the integrated pipeline against baseline and competing models under specified experimental conditions.
To address variable illumination and low-contrast defects, the authors extend the Adaptive Contrast Enhancement (ACE) approach by introducing an adaptive weight (AW) calculation that dynamically fuses color channels guided by information entropy. The AW-ACE method aims to normalize channel contributions and perform multiscale extraction and fusion, enhancing defect-relevant features across a variety of lighting conditions without manual adjustment for specific defect types. The adaptive fusion is intended to preserve subtle defect cues that can be lost under conventional enhancement or fixed-parameter methods.
Data scarcity and class imbalance—especially for small or low-contrast defects—motivate a generative augmentation strategy. The authors modify a DCGAN architecture to create ES-DCGAN by integrating an Efficient Channel Attention (ECA) mechanism into both the generator and discriminator and adopting the SELU activation function. These changes are designed to steer the generator toward producing more diverse and realistic defect samples and to reduce mode collapse by emphasizing defect-relevant channels and stabilizing feature distributions. ES-DCGAN-generated images are intended to enrich training diversity and to improve detector robustness to hard-to-capture defect variations.
The detection stage centers on a lightweight variant named GE-YOLO11n. The authors redesign the backbone and detection head to incorporate GhostConv operations and the ECA module to reduce parameter count and computational cost while improving the representation of small-scale defects. The architecture emphasizes lightweight feature fusion and multiscale sensitivity to retain fine-grained spatial information commonly required to detect small or low-contrast defects. The detector is evaluated for both accuracy (mAP) and latency, with attention to detector-only and full-pipeline timings relevant to real-time industrial inspection.
Key quantitative outcomes reported by the authors include:
The authors also report that the accessible data and code supporting the study are available on Zenodo; readers should consult the repository for dataset specifics, training hyperparameters, and a complete breakdown of comparative baselines and metrics.
The paper presents a coordinated, efficiency-minded solution for industrial steel plate surface defect detection. By combining an adaptive enhancement algorithm (AW-ACE), an attention-augmented generative augmentation (ES-DCGAN), and a lightweight attention-aware detector (GE-YOLO11n), the authors report improved detection accuracy and reduced computational cost suitable for real-time quality control. The complete pipeline delivered reported gains in mAP, a decrease in model parameters, and low full-pipeline latency. The authors provide data and code on Zenodo to support reproducibility and potential industrial adoption.