Enhanced Face Sketch Synthesis Through a Multistage GAN Framework with Hierarchical Discriminators and MobileNet Feature Extraction

Facial sketch synthesis is a crucial task in computer vision with applications in law enforcement, digital art, and human–computer interaction. Despite its importance, generating high-quality and realistic facial sketches from images remains challenging due to the difficulty in preserving fine-grained details and minimizing artifacts like blur. Existing methods often struggle to maintain a balance between global coherence and local detail enhancement, resulting in suboptimal sketch quality. In order to tackle these issues, we suggest a novel multistage Generative Adversarial Network (GAN) framework incorporating hierarchical discriminators and MobileNet-based feature extraction. The input facial image is divided into multiple local regions, such as the eyes, nose, and mouth, enabling the model to learn detailed features for each facial component. The framework is composed of three key modules. The Local Region Module synthesizes sketches for specific facial regions, ensuring accurate detail representation. The Mask Module utilizes a MobileNet architecture enhanced with Squeeze-and-Excitation (SE) blocks to extract robust features. It also calculates two custom loss functions, Mask_Structure (MS) loss and Mask_Feature (MF) loss, to improve sketch precision and detail preservation. The Adaptive Grasshopper Optimization (AGHO) algorithm is applied to fine-tune the parameters of the MobileNet model, thereby enhancing its performance and accuracy. Finally, the Fusion Module integrates outputs from local modules into a unified, realistic sketch. The proposed method achieved an accuracy of 96.82% on the CUHK dataset, 97.12% on the AR dataset, 98.84% on the XM2VTS dataset, and 98.62% on the CUFSF dataset, demonstrating its effectiveness across diverse benchmarks.