Artificial Intelligence Techniques enabled insights into leather defects

Abstract

With the advent of digital revolution, detection of leather surface defect gains immense significance towards automation in the assessment of leather quality which is of paramount importance in the leather trade that eventually happens to be global. The proposed work strives to develop an artificial intelligence enabled reliable and efficient system in detecting leather surface defects, using leather image dataset. The work uses conventional machine learning algorithms and deep learning approaches for distinguishing leather surfaces. However, it was later found that due to the variability in the leather surface and defects, the conventional machine learning algorithms were not able to satisfactorily distinguish the leather surfaces. LeatherNet, a novel lightweight deep neural network was proposed as a result. For better analysis, the performance of LeatherNet was compared with the performances of prominent existing convolution neural network models, already experimented machine learning algorithms and existing state of the arts in this domain. The performance of the LeatherNet was found to outperform all the algorithms, architectures and existing state of the arts taken into consideration. Accuracy, loss, precision, recall and AUC score metrics were used for performance measurement. When trained for 1500 epochs, the proposed model recorded a maximum training accuracy, precision, recall of 99.78%, 99.69% and 99.92% respectively, while the maximum testing accuracy, precision and recall recorded 97.42%, 97.66% and 99.40% respectively.