Abstract: In order to improve intelligent recognition accuracy and discrimination efficiency of tobacco diseases, improved disease recognition models based on the YOLOv5 network are proposed, aiming to improve the original model by addressing the weak extraction ability for small target spots. The improved models include YOLOv5-ME with the addition of a multi-scale enhancement module, YOLOv5-LT with the addition of a small-target detection layer, and YOLOv5-ME-LT with the addition of a small-target detection layer and a multi-scale enhancement module simultaneously. The YOLOv5-ME-LT model with both a small target detection layer and a multi-scale enhancement module were used to recognize and validate 5874 disease images collected from the field, including brown spot disease, cucumber mosaic virus disease, tobacco mosaic virus disease, tobacco weather fleck, and wildfire disease. The results show that the detection accuracies of three improved models are better than the original models, with 88.7% mAP for YOLOv5-ME; 88.1% mAP for YOLOv5-LT; and 91% mAP for YOLOv5-ME-LT, which are much higher than 78% of the original model. The improved tobacco leaf disease recognition algorithm based on this study has significantly improved the performance relative to the original model, but there are still problems such as the difficulty of distinguishing similar diseases. The follow-up study will combine life chemical indicators and multispectral technology to detect early tobacco diseases.