The use of CNNs for fruit recognition has increased significantly and has shown good results. This CNN is used to support a system that classifies ripe fruit during distribution. Five CNN models (MobileNetV2, InceptionV3, ResNet152V2, ResNet50V2 and Xception) are used to classify ripening fruit plums. We observed the performance changes of 5 CNN models by dropping out (applied, not applied) to a model that imported only the basic model and a model to which two types of transfer learning (freeze and relearning) were applied. We compared and analyzed which model was most suitable for the classification of ripening fruit plums under six conditions. As a result, the ResNet50V2 model, which was not applied in the case of dropout and re-learned in the case of transfer learning, showed a difference of 1 to 3% compared to other models using transfer learning. Based on this, it was concluded that using relearning of transfer learning without using Dropout in the problem of classifying plums can be more effective than using Dropout and freezing of transfer learning.