The Ronchi test is a simple and powerful for evaluating defects on the surface of a optical component, especially in concave mirrors. Consists in placing a grid with vertical, parallel, spaced lines uniformly and alternate between transparent and opaque, crossed by a beam of light that is reflected in the mirror and generates an interference pattern of light and dark fringes (Ronchigrams), which show the deviations of the front of wave with respect to the ideal. The identification and differentiation of the types of optical defects from the Ronchigrama is important for the quality evaluation of mirrors for telescopes that are traditionally carried out by a qualified operator, which is tedious and causes long-term vision problems. In this work, a learning-based approach is presented deep for automatic classification of optical defects of mirrors from rhonchigrams from a database of images acquired at the University Optics Workshop of the Plains. Different methods were used to increase data (i.e. spin, noise, lighting) and neural networks convolutional (CNN) as part of an integrated hardware and software prototype for automation of the Ronchi test. Preliminary results obtained with two different CNN architectures reached a performance of 98%, 98% and 97.9% for CNN1, and 98%, 96.9% and 98.5% for CNN2, for accuracy measurements (accuracy), precision and recall. These results show that the proposed approach can be used to detect and identify optical defects in mirrors carved in the optics workshop with high precision.