Original title: Asymmetric CNN–ViT Residual Fusion for Industrial Surface Defect Detection
Chao Ding, Kun Zou, Yu Zhong, Yong Liu +1 · September 2026
Factories need to find surface defects on materials even when they lack large amounts of labeled data. Researchers built an approach called ACVRF to spot these flaws. It combines two kinds of vision models. A convolutional neural network, or CNN, scans the image first. It proposes candidate defect spots and extracts the main visual features. Then, a Vision Transformer looks at those exact same regions to add extra residual detail through a controlled mixing limit. During training, frozen expert models also score mistakes to weight the training samples. The researchers found that ACVRF located defects better than a CNN by itself on two test setups. However, the system beat its strongest simple fusion alternative by only a tiny amount. Running it also took more than twice as long as the CNN alone, showing that the modest boost costs significant computer processing time.
You are sorting apples on a conveyor belt and need to spot any that are bruised. You have only seen a few bruised apples before, so you rely on a quick glance that sometimes misses subtle bruises. You decide to double-check with a magnifying glass only on the apples you thought might be bruised, but you are careful not to overcorrect the first impression. In this paper, the CNN is the quick glance that finds suspicious regions, and the Vision Transformer is the magnifying glass that gently adjusts the decision, but only for those regions.
Factories need automated inspection to catch tiny scratches or dents on surfaces, but they often have very few example defects to train on. This method improves detection accuracy by combining two different types of neural networks, which means fewer flawed products slip through. The downside is that it takes more time per image, so factories must weigh speed against precision.
On the GC10 dataset the evidence for a Transformer-specific contribution is limited. Using frozen expert losses for sample weighting during training slightly increased mean mAP50 but reduced mean mAP50:95 on both protocols. The model’s inference time (43.787 ms per image) is 2.4 times that of the CNN baseline (17.936 ms), and gains over the strongest concatenation baseline are only 0.00458 mAP50.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Asymmetric CNN–ViT Residual Fusion for Industrial Surface Defect Detection — Chao Ding, Kun Zou, Yu Zhong et al. — Eng—Advances in Engineering — doi:10.3390/eng7090492
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper.