Failures of mechanical systems are strongly related with wear of interacting surfaces in machine elements. Hence, wear monitoring is fundamental to avoid energy and time losses, as well as to prevent definite failures on machines. Wear monitoring can be achieved by capturing worn surfaces images, in which mass losses are represented as non-uniform texture patterns. This work introduces a computational framework to characterize and predict mild or severe wear regimes by using gradient-based descriptors. The HoG and Daisy descriptors were used to codify wear morphologies of worn surfaces images. Once images were coded as gradient patterns, the corresponding descriptors were mapped to a previously trained Support Vector Machine (SVM), allowing to automatically associate a wear regime label. A set of Scanning Electron Microscopy (SEM) images of abrasion worn surfaces were used to validate this work. The proposed framework achieves accuracy results of 94% and 96% using the Hog and Daisy descriptors, respectively.