A C-ResNet-50 model was proposed to improve the accuracy of computer recognition of potato leaf diseases in natural backgrounds, in response to the low accuracy of the existing algorithms. Firstly, images of late blight, early blight, anthracnose and healthy leaves for potatoes were collected in the field, and data augmentation was conducted by simulating factors such as shooting angle and weather conditions to construct an experimental dataset. Secondly, by comparing deep learning models, ResNet-50 network was selected and improvements were proposed. A 3×3 convolutional layer and a 1×1 convolutional layer with a step size of 1 were introduced into the residual block to reduce the severe missing feature information in the main branch of the residual block. A new fully connected layer was introduced to conquer the problem of high similarity and difficult classification of potato leaf diseases. The ECA attention module was added to address the issue of the insufficient targeted attention capability in the backbone network. The results showed that the average accuracy of the C-RseNet-50 network for identifying potato leaf diseases reached 90.83%, which was 1.84 percentage points higher than that of the original model.