Improved Perceptual Loss for Sketch Image Domain.

Abstract

Traditional perceptual loss functions, primarily designed for photographic images, often perform poorly in the sketch domain due to significant differences in visual representation. To address this domain gap, we propose an improved perceptual loss specifically designed for sketch images. Our method fine-tunes a pre-trained VGG-16 model on the ImageNet-Sketch dataset while strategically replacing max-pooling layers with spatial and channel attention mechanisms. We comprehensively evaluate our approach across three dimensions: generation quality, sketch retrieval performance, and feature space organization. Experimental results demonstrate consistent improvements across all evaluation metrics, with our method achieving the best generation performance, over 10% improvement in sketch retrieval accuracy, and 6-fold improvement in class separability compared to baseline methods. The ablation studies confirm that both fine-tuning and attention mechanisms are essential components that work together effectively. Our domain-specific perceptual loss effectively bridges the gap between photographic and sketch domains, providing enhanced performance for various sketch-related computer vision applications, including generation, retrieval, and recognition.