New CNNs with Complex Structure Tensor Boost Periocular Recognition by 26%

A new study from Kevin Hernandez-Diaz, Josef Bigun, and Fernando Alonso-Fernandez provides strong evidence that standard CNNs inherently struggle to extract orientation features from images. To solve this, the researchers propose feeding the networks with Complex Structure Tensors (CST)—compact representations that encode both orientation and certainty. This explicit, front-loaded approach mimics how mammalian vision processes orientation cues. When tested on publicly available periocular (eye-region) datasets Cross-Eyed and PolyU, the CST-enhanced CNNs achieved a 5–26% reduction in Equal Error Rate (EER) for biometric identification and verification, consistently outperforming traditional grayscale inputs. The improvement held across all six tested CNN architectures and both open-world and closed-world scenarios, where unknown identities must be rejected.

Crucially, the upfront use of CST allowed the researchers to reduce CNN depth and complexity without sacrificing accuracy. Smaller, mini-complex convnets feeding into reduced-size CNNs actually outperformed full-fledged, prevailing architectures. This suggests that explicit orientation priors not only boost accuracy but also improve explainability and model efficiency—key for deployment on thin clients like smartphones or embedded devices. The work, posted on arXiv (2607.15410), reinforces that incorporating biologically inspired feature representations can overcome fundamental limitations in current deep learning approaches for image recognition.