Guavira is a Brazilian Cerrado's biome native fruit, from the Campomanesia spp. species, that generates income for small producers and indigenous people. Due to its seed's recalcitrant power, guavira is very difficult to produce on a large scale. This paper introduces the first computer vision proposal for the classification of guavira seeds viability from images captured after the tetrazolium salt test with mucilage and lime. The proposed approach makes use of SLIC (Simple Linear Iterative Clustering) to segment the images into superpixels which are later classified into viable or not-viable seeds by machine learning algorithms. Six different algorithms were compared: Multinomial Naive Bayes, AdaBoost, Random Forest, Multilayer Perceptron, k-Nearest Neighbors, and Support Vector Machines. To train the algorithms, 411 superpixels for mucilage and 1,307 for lime were classified by a specialist into viable and not viable. After a comparative analysis of the six algorithms, Multilayer Perceptron with the balancing method SMOTE obtained a better accuracy with 97.92% correct recognition for mucilage and 96.32% for lime. Given the results, we conclude that the presented method can be a useful tool in future studies and experiments on guavira seeds.