Randa Alsabhi
The proposed framework demonstrated strong performance, achieving an Intersection over Union (IoU) score of 0.93, an average landmark alignment error of 3.2 pixels, and an aesthetic consistency score of 9.5/10. These results outperformed the CNN and conditional GAN baseline models across the evaluated performance measures.
INTRODUCTION: Fashion pattern generation remains one of the most labor-intensive stages in garment production because it depends heavily on expert manual drafting, iterative revisions, and technical precision. Although existing artificial intelligence (AI)-based fashion systems have primarily focused on garment classification, image synthesis, and trend prediction, limited research has addressed the automated generation of structured garment pattern representations incorporating manufacturing-oriented construction information. This study addresses this gap by developing an end-to-end deep learning framework for automated fashion pattern generation.
METHODS: The proposed framework converts garment images, sketches, and textual construction descriptions into structured, CAD-compatible fashion pattern representations. It integrates Vision Transformer (ViT)-based visual feature extraction, BERT-based textual encoding, multimodal feature fusion, and an encoder-decoder Generative Adversarial Network (GAN) for iterative pattern generation and refinement. Automated seam and notch generation modules are additionally incorporated to enhance manufacturing compatibility and CAD integration. The framework was experimentally evaluated using the DeepFashion dataset and compared with CNN and conditional GAN baseline models under identical experimental conditions.
RESULTS: The proposed framework demonstrated strong performance, achieving an Intersection over Union (IoU) score of 0.93, an average landmark alignment error of 3.2 pixels, and an aesthetic consistency score of 9.5/10. These results outperformed the CNN and conditional GAN baseline models across the evaluated performance measures.
DISCUSSION: The findings demonstrate the potential of the proposed multimodal deep learning framework to provide a scalable and production-oriented solution for automated fashion pattern generation. By integrating visual, textual, and manufacturing-oriented information, the framework can reduce pattern development time, improve pattern consistency, and support mass customization within digital fashion production workflows.