CrimeSketch AI: Text-To-Image Synthesis for Criminal Identification Using Deep Learning

Abstract

Facial identification based on verbal eyewitness descriptions remains a significant challenge in contemporary forensics, particularly in the absence of visual evidence. Traditional methods such as CNN-based models and Haar cascade classifiers struggle with occlusion, poor lighting, and low-resolution inputs, while artist-drawn sketches are impractical for scalable automation. To overcome these limitations, we introduce CrimeSketch AI, a deep learning pipeline for generating realistic facial images from natural language descriptions. Our system combines Stable Diffusion fine-tuned by LoRA for text-to-sketch generation, Canny edge detection with ControlNet or CycleGAN for sketch-to-image translation, and StyleGAN2-ADA with CLIP for text-to-face synthesis. Identity matching is done using InsightFace with ArcFace embeddings to assure the accurate recognition. The model was tested on publicly released datasets, such as CelebA and a cleaned criminal mugshot dataset. CrimeSketch AI presents a scalable, privacy-preserving method for AI-powered suspect identification, augmenting the capability of law enforcers based on text-based descriptions alone.

More by Jyotsna Chandrasekharan

• Optimization of Lotka-Volterra Equations using Ant Colony OptimizationInternational Conference on Recent Advancement in Electrical, Computer and Communication Technologies · 2026
• Comparative Analysis of Deep Learning Models for Visual Classification of Terrestrial and Extraterrestrial Rock ImageIEEE International Conference on AI Engineering and Innovations · 2026 · Citations: 1
• Real Time Screen Based Gaze Estimation Using MediaPipe and Machine Learning ModelsIEEE International Conference on AI Engineering and Innovations · 2026
• ECG Analysis for Myocardial Infarction Detection via Deep LearningProcedia Computer Science · 2026