Automated Knitting Instruction Generation from Fabric Images Using Deep Learning

This paper proposes a deep learning method to generate knitting instructions from fabric images, reversing the traditional process of manual pattern coding. The instructions are based on the five-layer structure derived from knitting logic, consisting of a front label recognizable from the front real image and a complete label corresponding to the knitting programming. The approach is divided into two phases: the Generation Phase, where neural networks produce front label instructions, and the Inference Phase, where mesh-based logic expands these instructions into complete labels. We also introduce an optimized labeling method for Kaspar's dataset, improving the accuracy, scalability, and knittability of the generated instructions. This method simplifies the knitting process by automating instruction generation directly from fabric images.