Larese and Granitto essentially consider scale-invariant feature transform (SIFT) plus bag-of-words (BOW) plus support vector machines (SVM) to detect relevant vein patterns that characterize different species and varieties of legumes.
Though the paper is nicely written and technically very sound, its basic concept can be further extended for classifications of overlapped patterns. The most important concept is BOW for extracting numerical features for the classification of patterns. In this context, the authors mention a reference (number [9] in the paper, [1] here), but they do not demonstrate how BOW is derived for the class of problems they are considering. The most common way to extract numerical features from text content, namely tokenizing, counting, and normalizing, is called BOW. Thus, a corpus of documents can be converted into metrics with one row per document and one column per token (as per, for example, word) occurring in the corpus. Vectorizing this general process of turning a collection of text documents into numerical feature vectors is a very challenging task about which the authors do not mention anything except reference [9] in section 4.2. Thus, the entire concept of the paper, which is very impressive, can be further improved by considering such details for wider classes of problems of a similar nature.
Of course, the authors compare the performance of their algorithm with some of the existing works given in reference [15] in the paper ([2] here), but such a comparison does not really clarify the complexity of classification under overlapped conditions of the classes.
The overall presentation and reported results are quite satisfactory.