Wheat stripe rust and wheat yellow dwarf have posed a great threat to the yield and quality of wheat. Early identification of these two diseases has important implications for the prevention and control of wheat diseases. Drought, nutrient deficiency and bacterial disease can lead to chlorosis and yellowing of plant leaves. These phenotypic symptoms are similar to infected leaves of wheat stripe rust and wheat yellow dwarf. In addition, the infected leaves of these two diseases are similar to healthy leaves because of indistinct phenotypic symptoms in the early stage of diseases. It is difficult to quickly and accurately distinguish them by existing identification methods. In this study, an improved Faster Regions with CNN Features (Faster-RCNN) disease identification method was proposed. There are two improvement strategies in our proposed method. Firstly, in order to enhance the fine feature extraction capability of the entire network and reduce the number of model parameters, three 3×3 grouping convolutions and down-sampling delay were employed to optimize the Deep Residual Neural Network (ResNet-50), which was designed as the backbone feature extraction network. Secondly, ROI Align was employed instead of ROI pooling to reduce the feature error problem caused by double quantization. It is helpful to solve the difficult problem of distinguishing subtle differences. Meanwhile transfer learning was employed to improve the training speed of the model and data augmentation was utilized to reduce over-fitting problems, which can further enhance recognition performance and generalization ability of our method. Experiments were carried out on a self-built data set of disease leaf images covering more than 200 wheat varieties showing different resistance and susceptibility to diseases while covering various symptoms at different disease stages. Performance indicators such as loss function convergence curve and mean mean precision (mAP) were selected to evaluate the effectiveness of the improved strategy. The experimental results showed that the mAP of the improved Faster-RCNN identification method proposed in this paper was 9.26% higher than the SSD, 7.64% higher than the YOLO, and 14.97% higher than the Faster-RCNN. The mAP of our proposed method reached 98.74% for wheat stripe rust, wheat yellow dwarf, healthy wheat and wheat with other etiolation symptoms. Moreover, in order to predict the diseases as early as possible, the early identification of disease infection was strengthened in this study. Our dataset contains 683 and 630 mild symptom photos of these two diseases respectively. The mAP for mild and severe symptom identification of these two diseases reached 91.06% by utilizing our proposed method. In addition, the value of the loss function decreased faster, as well as model performed better overall. Finally, In order to implement the deployment and application of our proposed method, the intelligent recognition system of wheat disease was developed, and WeChat applet was used to identify wheat diseases in the field. Under the condition of maximum concurrent access of 100, the average return delay was 5.024 seconds, and the success rate of recognition return was 97.85%, and the average accuracy of the recognition of two kinds of wheat diseases and their subdivision was 93.56%. The system can effectively meet the practical application requirements and be employed to guide the scientific prevention and control of diseases.