Abstract:
Fibers are usually twisted together to form yarns, and yarns are then woven or knit to form fabrics.
All textiles are made of fibers, but not all fibers can be used to make textiles. Besides, fibers are
classified into two major categories according to their origin that is manmade and natural, each
major category has also a group. The knowledge of identifying the textile fibers helps a producer
of garments forensic and designers to identify the type of fiber and the care to be taken in
maintaining the fabrics made of a particular type of fiber. This is an important factor for labeling
the garments, which includes specifying the fiber content in the garment. The application of
imaging techniques on fiber classification had been studied but the number of fiber types are
limited and all applied a single approach. This single approach will lead to misclassification and
decrease the accuracy. Even if the latest microscopic identification approaches are also dependent
on human beings' skill and eyes. The necked eye observation is not perfect and also the detection
by the experts will be time-consuming and tiresome. Another important concept is real time
detection of textile fibers in the manufacturing process hence quality is one of the determinant
factors in textile industry.
In this study, we apply a GLCM assisted CNN for textile fiber classification. We collect 746
microscopic image datasets of cotton, wool and towel from Ethiopian Institute of Textile and
Fashion Technology [EiTEX] and cameo online textile fiber image library. After collecting the
dataset, preprocessing like interpolation, noise reduction and segmentation applied to get an
enhanced image. A CNN trained model is developed for classification and extraction. Our model
has a concatenating with residual module to get a multi-level feature and for skipping operation
that has a gap in sequential layers.
As a backend python with Keras and Tensorflows tools are implemented in our model. Our models
are evaluated using a metrics accuracy, precision, F1-score and recall. We test our model with
different optimizers to determine the worst and best scores. The planned model achieves 99% and
77% with Adam and Adadelta optimizers respectively with batch sizes of 32 and 100 epochs. Our
model sounds good with kernel regularizer and using GLCM as an input to the CNN parameter.
The study has also a limitation of classifying a poorly preprocessed image dataset and
classification of only cotton, wool, and towel prominent textile fiber types.
Keywords: Preprocessing, Optimizers, Cotton, Wool, Towel, CNN, GLCM