There is a need for an accurate estimate of the percentage of vehicle damage to make insurance claims and assess repair costs. The given paper applies three widespread deep learning architectures, including AlexNet, VGG16, and ResNet, to the problem of the degree of the damage of vehicles on an image. Damage may be classified as a percentage and, therefore, more numerically represented; the solution by a simple image classification technique may be used. In the case, the authors tried to apply a dataset of vehicle images with different degrees of damage, and each image was accurately annotated by the corresponding percentage of the damage. The approach to the implementation for each model separately included the use of model training by transfer learning with the used pre-trained weights on a large dataset to boost the performance and decrease the model training time. Also, data augmentation was applied to increase the model performance across a broader range of damage properties. The results reveal that the performance of the ResNet architecture is better for both others. It is possible to classify from the images the damage percentage with the help of the VGG16 architecture. The major reason this architecture outperforms others is that it can learn a higher degree of complex features with the help of residual connections. The main contribution of the exploration is the development of more sophisticated algorithms to boost the percent of vehicle damage recognition.