The chapter presents a comparative analysis of three types of classical deep neural network methods used for classification of color images of corn seeds from 17 varieties, previously divided into two classes—sound and infected with the disease Fusarium Moniliforme. The SqeezeNet, GoogLeNet, and ResNet50 deep learning methods were used to determine the variety and the condition of the captured images. The sets were divided both as 34 classes and as 2 classes problem solving. The comparison between the achieved classification accuracy with three CNNs on the test sets is made by different evaluation metrics and is analyzed. It shows that in terms of classification accuracy in 34 classes the ResNet50 had highest value of mPA0.5 = 76.09%, while in two classes GoogLeNet achieved the best accuracy of 97.54%, and obtains only 0.72% difference between test and validation sets. The probability of belonging to classes has been evaluated also.