Classifying the severity of crop diseases is the staple basic element of the plant pathology for making disease prevent and control strategies. In order to achieve better results in the classification of the severity (healthy, general or severe) of crop diseases, a CDCNNv2 algorithm based on residual network (ResNet 50) and deep transfer learning was proposed. By training more than 30,000 crop disease images which were divided into 10 species, a model for the classification of disease severity was obtained, and the recognition accuracy could reach 91.51%. For verifying the robustness of the CDCNNv2 model, comparative experiments were carried out with ResNet 50, Xception, VGG16, VGG19 and DenseNet 121 that used transfer learning. The experimental results showed that the average accuracy of the CDCNNv2 model was increased by 2.78~10.93 percentage points, which had higher classification accuracy and strengthened the robustness of crop disease severity identification. At the same time, based on the model trained by this algorithm, combined with Android technology, a realtime and online crop diseases severity recognition APP was developed. By photographing the disease areas of the crop leaves, the recognition results (species-disease-severity) and recommendations for prevention and treatment can be obtained between 0.1s and 0.5s. In addition, other related supporting functions such as disease encyclopedia made the APP more complete, which can provide effective solutions and ideas for the prevention and treatment of crop diseases.