High-Performance Halftoning with Binary Loss Optimization and Efficient Gradient Descent

This paper presents a novel halftoning algorithm based on continuous relaxation and gradient-based optimization. By reformulating the original binary quadratic programming (BQP) problem into a differentiable non-convex objective, the proposed method integrates a perceptual loss with a binarization regularizer to guide the optimization toward high-quality binary outputs. Unlike deep learning-based approaches, our method is model-free, memory-efficient, and fully interpretable, enabling fast convergence without complex training or large memory overhead. Experimental results demonstrate that the proposed method outperforms both classical and neural halftoning techniques in terms of image quality and convergence speed. This work provides a practical and scalable solution for highresolution halftone image generation, particularly suitable for industrial and real-time applications under resource constraints.