Automatic mustache pattern production on denim fabric with generative adversarial networks

This thesis covers current approaches and proposed new architectures on the automatic detection and production of "mustache" patterns in denim fabrics. The problem in the focus of the thesis is the automatic and high quality production of mustache patterns in denim images. This study was determined as a result of interviews with Baykan Denim company operating in Malatya 2nd Organized Industrial Zone. Company customers demand similar mustache-patterned fabrics by bringing a sample mustache-pattern denim fabric. Classic mustache pattern production is done by Photoshop experts working on the sample for approximately 2-3 hours. High personnel cost, long duration of the process and subjectivity (experts can produce different mustache patterns from the same image) are the disadvantages encountered in the field. A new dataset consisting of high-resolution denim-mustache images has been constructed in order to avoid the aforementioned negativities and to produce suitable solutions for the denim-mustache translation problem. With this dataset, architectures that provide generative network-based translation are comparatively examined. Inspired by the examined architectures and considering their disadvantages, original approaches have been brought to the literature.