Improvement of phenotyping in legume breeding is a major challenge to increase yield or to imagine novel food uses. The analysis of large genetic resources collections requires for developing fast, cheap and reliable screening tools and related data processing. Artificial vision with several lighting of different wavelengths seems to be a good alternative to characterize seeds of genetic resources collections. This technique consists in the acquisition of a set of spectral images on a unique sample. The aim of this work is to develop a new methodology of seeds phenotyping, which can be applied to different legume species. To achieve this goal, we have used multispectral imaging and chemometrics for evaluating the phenotypic variability of the internal structure of faba bean and lupin seeds. For each cultivar, 10 seeds were cross-sectioned at half-length and a multispectral image (spatial size of 10 mm * 9 mm) of this section was acquired. The images gave many relevant pieces of information about the variability of some criteria related to internal texture, shape parameters and color. Some spectral signatures were assigned to tissues and exploited to label histological areas in seeds. PCA analysis of these images highlighted that some cultivars a low intra-genotype variability, and that the inter-genotypic variability was higher than intra genotype for shape of seeds.