This paper compares the efficiency and accuracy of ResNet-34 and DenseNet-BC-121 neural networks for recognizing 50 types of fruits using a dataset of 25,100 training images and 12,700 testing images. The results indicate that ResNet-34 outperforms DenseNet-BC-121 in terms of training efficiency, achieving over 98% accuracy significantly faster. The findings suggest that the architecture of ResNet, which utilizes residual connections to mitigate gradient issues, leads to better performance on simpler datasets compared to DenseNet.