The classification of physical defects, with the aim of ensure the quality of arabica green coffee beans, is important from a commercial point of view. This classification is done mostly by human experts, which are slow and error prone. The main works in the literature focused on solving the problem using computer vision, require prototypes, which take each image from a completely vertical angle to the surface where the sample is coffee beans. Each of these prototypes is a limiting practical work, because of their difficulty of implementation and the restrictive angle. Seeking a solution to these problems, an automatic sorter twelve physical defects is presented, using images acquired by a cell phone with an angle of diagonal shot, similar to that made by taking a picture of an object located at a lower altitude normally. The classification results show a 100% overall accuracy.