Facial Marks as Biometric Indicators for Celebrity Prediction

The widespread availability of celebrity images and videos online has driven significant advancements in training deep neural networks for celebrity face recognition. While deep learning excels in analyzing complex visual attributes such as identity, age, gender, and race, the internal mechanisms governing these models remain largely opaque. In contrast, human visual processing is guided by well-established biological principles, utilizing saccadic eye movements to efficiently gather detailed information from salient regions in the visual field. This process leverages the retina's varied resolution and retino-cortical mapping, enabling simultaneous broad scene comprehension and precise detail capture. Inspired by these principles, this paper introduces a novel face analysis system that integrates the HMAX model―a biologically-plausible, hierarchical visual architecture―with a log-polar image representation, which simulates early-stage visual processing in primates. The proposed system employs a space-variant sampling method that enhances context awareness by combining peripheral, low-resolution visual information with high-resolution details from central visual regions. This design enables effective identification and analysis of both coarse structural features (e.g., face shape) and fine-grained details (e.g., scars, moles, wrinkles) critical for robust face recognition. By bridging biological insights with state-of-the-art computational techniques, the proposed approach offers a promising path toward more interpretable and efficient visual analysis models.