Real-Time Gumbel-Softmax Adaptive Thresholding for Robust Object Classification under High Illumination Conditions

This paper proposes a novel object classification system integrating a dual-imaging CMOS Image Sensor (CIS) with a real-time Gumbel-Softmax-based threshold selection module to address feature loss under high illumination conditions. The proposed system reliably extracts binary feature information from saturated regions directly at the sensor stage and implements a dynamically learnable thresholding approach. Experimental results using the ImageNet dataset demonstrated that our method achieves higher structural similarity (SSIM) and classification accuracy compared to conventional thresholding techniques, verifying robust classification performance across varying illumination scenarios. This research provides clear insights into future directions for next-generation vision system development through the integration of hardware and real-time deep learning methodologies.