DEEP LEARNING MODEL FOR LONG FIBER, SMOOTHNESS, AND TOWELING EFFECT RECOGNITION ON THE COMMONS DATASET

• Main Author: Lafifa Zulfa, Iqbal
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/81136
• Institution: Institut Teknologi Bandung

The deep learning approach to representing the characteristic textures on the surface of fabrics offers a solution to replace the subjective measurements previously conducted by textile experts. This objective representation plays a crucial role in various industries, such as textile material recognition, product design, and quality control. Ins this study, we develop a deep learning model capable of providing highly accurate texture representation with optimal time efficiency. The deep learning model was trained on the CoMMonS dataset, comprising 6912 images of 24 textile materials. The data were analyzed and classified by experts into three categories, namely fiber length, smoothness, and toweling effect, serves as a significant benchmark for deep learning models. We have selected to use VGG-S-BN (VGG-S with Batch Normalization) as the primary framework, known for its effectiveness in image processing and also the model will become a solution for computational cost problem. The results show that VGG-S-BN outperforms VGG16 and MuLTER in terms of computational cost, accuracy, and loss value in extracting features from the dataset. The accuracy values obtained by VGG-S-BN are 67.36%, 55.13%, 67.36% on the training data and 74.32%, 65.23%, 74.32% on the validation data with the best learning rate set at 0.001. The research findings are expected to make a significant contribution to improving the accuracy and efficiency of fabric texture representation in the textile industry.