Computer vision (CV) techniques help to perform non-destructive seed viability detection (SVD) for faster, more efficient and fairer results. However, the seed vigor dataset currently suffers from insufficient number of samples, data noise, and imbalance of positive and negative samples.