Neural networks, deep learning, NLP · insight-1472

Probe a CNN with texture and capture shifts before moving it to a new camera or catalogue

Situation
A product-image or defect CNN trained under one camera and lighting setup is moved to another source, such as customer photos instead of catalogue shots.
Action
Build a small shifted evaluation set (blur, noise, grayscale, different camera, texture or style changes), report accuracy per shift against the original, and add the shifts as training augmentations if the drop matters.
Reason
Geirhos et al. found that ImageNet-trained CNNs rely strongly on texture over shape on cue-conflict images, and that training on texture-altered images increased shape bias and robustness to distortions.
Boundary
Transfer-learned image models deployed to a new capture setting.

Symptom or risk

ImageNet-trained CNNs lean on texture, so changes in surface texture, noise or filtering can flip predictions even though the object is unchanged.

Tasks where texture is the true signal, such as fabric defects.

Shift sets must be built and kept; augmentation can cost clean accuracy.

Calling any accuracy drop a "texture bias" without testing; using only synthetic shifts that never occur.

Geirhos et al. report that ImageNet-trained CNNs rely strongly on texture on images with conflicting cues, and that Stylized-ImageNet training increases shape bias and robustness to distortions.
ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness