Image-recognition AI can ace clean test images yet stumble on blurry or noisy ones. DRAGS is a simple training trick that tells the model to stop learning from the parts of an image that look like noise.
The problem
Deep vision models often break down when inputs are corrupted by noise, blur or occlusion. Common regularisers act uniformly across the whole image, rather than focusing on the problem areas.
What the research did
We introduced Dynamic Region-Aware Gradient Suppression (DRAGS). During training, it scores each location in a feature map, keeps only the most informative fraction, and stops learning signals from the rest, which are likely noisy or misleading.
What we found
On CIFAR-10 with a ResNet-18 model, DRAGS improved clean accuracy by 1.39 points over a strong baseline (64.15% ± 2.30 versus 62.76% ± 1.05 across three seeds) while maintaining robustness under common corruptions, with slight gains on Gaussian noise.
Who benefits
- Engineers deploying vision AI in messy real-world conditions
- Researchers in robust and trustworthy AI
- Applications such as medical imaging and autonomous systems
Limits and what comes next
Gains are modest and tested on one dataset and model; the variance across seeds means larger studies are needed. This connects to my PhD focus on trustworthy AI.
The team
This research was led by Md Muntaqim Meherab with me and colleagues from Daffodil International University, South Dakota School of Mines and Technology, Islamic University of Technology and BUET.
Frequently asked questions
What does DRAGS do?
It stops a vision model learning from image regions that look noisy or misleading during training.
How much does it help?
About 1.39 points of clean accuracy on CIFAR-10 in the study.
Why does robustness matter?
Real-world images are often blurry or noisy, and AI must still work reliably.