Um Sistema de visão computacional para classificação da qualidade do couro caprino

Abstract

An important economic alternative for the semi-arid region of Brazil is the goat/sheep farming. Besides milk and meat, goat/sheep skins are much appreciated in the manufacturing of fine artifacts (e.g. shoes, bags & purses, wall ets, and jackets). However, due to the extensive mode of raising/breeding and the informality of slaughtering, sheep/goat farmers deliver to industry skin pieces with different types and levels of defects. Then, at the industry, specialized workers have to classify/discriminate the skin pieces according to their qualities. This handmade work is time - consuming and extremely dependent on the experience of the employee in charge of the skin - quality discrimination. Even the same employee may produce different classifications if he/she is asked to reclassify the skin lot. Thus, in order to handle these problems, in this paper we report the first results of a computer vision based system aiming at classifying automatically the quality of goat/sheep skin pieces. For this purpose, we compare the performances of statistica l and neural network classifiers using several feature extraction techniques, such as Column - Variance (VAR), Haar wavelet transform (HAAR), Non - Negative Matrix Factorization (NMF), Principal Component Analysis (PCA) and Gray Level Co - occurence Matrices (GL CM). We also implemented the reject option in the evaluated classifiers. Reject option is a technique used do improve classifier's reliability in decision support systems. It consists in withholding the automatic classification of an item, if the decision is considered not sufficiently reliable. The rejected item is then handled by a different classifier or by a human expert. By means of an in - depth analysis of the results, it was possible to observe that, without the reject option mechanism, the VAR, NMF a nd HAAR techniques achieved the best performances when associated with the ELM and SVM classifiers. When the reject option mechanism was present, it was observed a considerable improvement of the classification rates, at the expenses of relatively high rej ection rates. It was also possible to observe that, for the evaluated classifiers, the HAAR and GLCM techniques were less affected by the use of the reject option mechanism in comparison to the results achieved for the case without reject option