Agricultural product recognition system using taxonomist's knowledge as semantic attributes

Abstract

Support Vector Machine (SVM) was used to classify type of produce commonly sold in supermarkets. We applied a sequence of image processing algorithms such as conversion of color space, thresholding and morphological operation to obtain the region of interest from the images. Global and local features were extracted from the images and used as input for the classifiers. The color and texture features extracted in this system were L*a*b* values and texton approach respectively. Since attribute learning has emerged as a promising paradigm for assisting in object recognition, we proposed to integrate it into our system. This could tackle problem occurred when less training data are available, i.e. less than 20 samples per class. The performances of the proposed classifier and conventional SVM were also compared. The experiments showed that the classification accuracy of the proposed classifier is higher than conventional SVM by 7% when only 4 samples per class were trained.