Abstract: There are many kinds of tobacco leaf diseases with complex pathology, which seriously affect yield and quality of tobacco. Accurate detection of tobacco diseases is prerequisite for timely prevention and control of tobacco diseases. Traditional detection methods have poor accuracy and low efficiency, and algorithms based on deep learning can improve accuracy of tobacco disease detection. In this study, five common tobacco diseases (common mosaic disease, cucumber mosaic virus disease, scab disease, tobacco wildfire disease, and climatic spot disease) were taken as research objects, and a tobacco disease detection model based on YOLOv3 was constructed to achieve accurate and rapid detection a variety of tobacco diseases. Darknet53 feature network was used to extract tobacco leaf disease features and fuse disease features of different scales. K-means++ algorithm was used to classify and position the fused features, and non-maximum suppression algorithm (NMS) was used to remove redundant frames to obtain disease area predictions frame. Using the tobacco disease data set collected in the field, the constructed YOLOv3 disease detection model and the SSD (Single Shot multibox Detector) model were compared and tested. The results showed that YOLOv3's mIoU of 0.81 was significantly better than SSD's 0.73, and the YOLOv3 model's mAP of 0.77 was also higher than SSD's 0.69. The YOLOv3 tobacco disease detection model constructed in this study can effectively locate the disease area of tobacco leaves, realize the detection of various types of tobacco diseases, and provide a reference for precise disease control.