Abstract: Buckwheat is widely cultivated in the regions of high altitude and cold mountains, such as northern and southwest China. The occurrence of crop diseases has posed a great threat to the quality and yield of buckwheat. Disease surveillance is an important measure to ensure the healthy development of the buckwheat industry. Since artificial intelligence has been extending to precision agriculture in recent years, machine learning and pattern recognition are beneficial to image classification, detection, and recognition for high accuracy and efficiency, while reducing the overhead in the detection of crop diseases. However, the existing deep learning cannot consider the complexity of disease images collected in the field, such as leaf overlap, uneven lighting, and shadow coverage. Therefore, it is necessary to accurately extract key features in a complex imaging environment for accurate recognition of crop diseases. In this study, a feasible multi-layer feature extraction in deep learning was proposed to extract the features of buckwheat diseases. A surveillance system of disease was then established according to the classified characteristics, thereby automatically identifying the categories of buckwheat diseases. Firstly, a combination of Maximally Stable Extremal Regions (MSER) and Convolutional Neural Network (CNN) was used to detect the feature regions of buckwheat disease, where the disease and disease-free areas were separated to precisely locate the disease position. A two-level inception structure was then added to the traditional CNN. The first layer was used to extract the contour features of the disease, where the dimension of the feature was represented by shallow information. The second layer was selected to obtain more detailed features, so that the number of parameters was relatively small in each layer, avoiding the gradient disappearance in the training process. This structure was normally used to extract the features for the low-quality images of buckwheat disease, particularly for the higher accuracy of classification. An improved convolution based on cosine similarity was utilized rather than the traditional convolution operation, in order to reduce the sensitivity of illumination during the sampling process. After that, the positions with similar features to the convolution kernel behaved higher activation values in the feature map. In addition, the difference between features was also reduced to prevent the interference of sample noise, thereby achieving better feature extraction for samples with uneven illumination. Finally, a systematic dataset was established, including some images for eight types of buckwheat diseases. The results showed that the combined framework of MSER and CNN with the two-level inception recognition was an effective way for region detection. In the identification for the presence or absence of disease, the accuracy, recall, and F1-measure reached 97.54%, 96.38%, and 97.82%, respectively. Furthermore, the mean for the recognition of disease categories reached 84.86%, 85.78%, and 85.4%, respectively, indicating excellent performance in recognition accuracy and speed. The finding can provide promising technical support for the automatic recognition of buckwheat diseases.