Handwritten Feature Descriptor Methods Applied to Fruit Classification

Abstract

Several works have presented distinct ways to compute feature descriptor from different applications and domains. A main issue in Computer Vision systems is how to choose the best descriptor for specific domains. Usually, Computer Vision experts try several combination of descriptor until reach a good result of classification, clustering or retrieving – for instance, the best descriptor is that capable of discriminating the dataset images and reach high correct classification rates. In this paper, we used feature descriptors commonly applied in handwritten images to improve the image classification from fruit datasets. We present distinct combinations of Zoning and Character-Edge Distance methods to generate feature descriptor from fruits. The combination of these two descriptor with Discrete Fourier Transform led us to a new approach for acquire features from fruit images. In the experiments, the new approaches are compared with the main descriptors presented in the literature and our best approach of feature descriptors reaches a correct classification rate of 97.5%. Additionally, we also show how to perform a detailed inspection in feature spaces through an image visualization technique based on a similarity trees known as Neigbor Joining (NJ).