Fabric Defect Detection using Deep Learning

Abstract

Quality control is a very important step in textile industry. While it is important to find out the defects in the fabric, it is also important to identify the location of the defect and to classify which type of defect it this, so that necessary steps in the production process can be taken to reduce the flaws. Apart from the manual labor work required for this method which is costly, time-consuming and error-prone, methods like gray-level co-occurrence matrix, Gabor filter, CZI-net have been previously used to identify imperfections in the fabric but due to their high processing time on large scale data, there is need for some other defect detection method that is faster, takes less manual work to be done and processes large scale data efficiently. With the help of new advances in the field of Artificial Intelligent (AI) and Deep Learning (DL), such highly efficient algorithms can be used that are fit for the modern day needs. Which not only reduces the processing time, but are also scalable as the industry grows and are also highly accurate. This research shows how Deep Learning algorithms like MobileNet can be used, which are smaller in size so takes up less space on the system and low processing time is consumed. ResNet and VGG-16 are also used which are bigger algorithms and are more complex in architecture. The results from the algorithms are further optimized using Bayesian Optimizer that sets the best hyperparameter combination of the algorithms with which optimal results can be achieved. Also, these state-of-the art algorithms have depicted promising results on large scale dataset.