Research on dynamic simulation and optimization of garment wrinkles combining computer vision and image processing

• Proposes a novel hybrid framework combining RGBD-based CNN feature extraction, LSTM temporal modeling, and reinforcement learning for real-time wrinkle detection and correction. • Integrates physics-based mass–spring simulation with adaptive stiffness and damping control to enhance fabric realism under dynamic conditions. • Achieves high classification accuracy of 96.4 % across shallow, medium, and deep wrinkles, with significant gains over 2D models. • Demonstrates robust wrinkle reconstruction under occlusion using Mask R-CNN–based segmentation and synthetic occlusion training. • Validates real-time performance at 30 FPS with low temporal drift (0.0187) and MSE (0.0246), enabling applications in virtual fashion, robotics, and AR/VR environments. This study presents a novel framework for real-time cloth wrinkle detection and optimisation, combining physics-based modelling with LSTM-Reinforcement Learning (LSTM-RL) and advanced computer vision techniques. A curated dataset of 45,876 annotated static garment images was used, featuring wrinkle attributes such as location, depth, width, and geometry. CNNs were employed for feature extraction, enhanced by Mask R-CNN to handle occlusions and RGBD data for depth-aware wrinkle modelling. A mass-spring system simulated fabric dynamics under environmental forces, while LSTM networks predicted the temporal evolution of wrinkles. Reinforcement learning dynamically adjusted fabric parameters, improving adaptability. The proposed RGBD model achieved a wrinkle detection accuracy of 96.4 %, outperforming the 2D model by over 9 %. Key metrics include an MSE of 0.0246, drift of 0.0187, and a reward value of -0.13543, with low policy and value loss confirming the RL agent’s learning stability. These results demonstrate high accuracy, real-time performance, and robustness, with strong implications for virtual fashion, robotics, and AR/VR applications.