PERTANIKA JOURNAL OF SCIENCE AND TECHNOLOGY

The synthesis of forensic face sketches is a crucial component of law enforcement, assisting in suspect identification based on eyewitness accounts. Conventional approaches, such as relying on forensic artists or composite sketch software, often suffer from subjectivity and inefficiency, leading to inconsistencies in quality and accuracy. This research introduces a novel method leveraging a dual-stage Generative Adversarial Network (GAN) architecture, conditioned on textual descriptions, to automate the forensic sketch generation process. The first stage produces a preliminary sketch that establishes the foundational facial structure, while the second stage enhances the sketch with intricate details like facial hair and accessories. Additionally, an adaptive stop training mechanism is implemented to terminate training when the generator and discriminator exhibit stagnation, thereby optimising computational efficiency. By incorporating GloVe and LSTM embeddings for encoding textual descriptions, our model effectively interprets complex linguistic inputs. The proposed framework is assessed on forensic sketch datasets, demonstrating superior performance over traditional techniques both qualitatively and quantitatively. This approach not only streamlines forensic sketch creation but also improves accuracy and realism, positioning it as a valuable asset in criminal investigations.