Fault Classification on Melamine Faced Panels Using Local Binary Pattern

The wood-based industry is the focus of users that require changes towards a clean industry, environmentally friendly and with efficient use of natural resources. Tasks of inspection and quality control are essential in this scenario. In this work, a dataset with samples obtained from near-infrared (NIR) image acquisition is used to evaluate the limits of the local binary pattern (LBP) for quality control of melamine board products. Conventional pattern recognition and convolutional neural network (CNN) approaches are compared concerning their use to classify the most common groups of faults present on the plant for the inspection task. The local binary convolutional neural networks (LBCNN) is used for inspecting, in a CNN inspired by the traditional LBP texture descriptor. The work shows that such a reformulation of the standard LBP is very simple and enables similar results. However, the results present better performance when LBP is combined with another type of feature, even only based on intensity. Similar modifications of standard CNN can be tested to promote the development of new CNN models insensible to texture granularity, image resolution, intensity range, and other variations of the acquired samples.