Panax notoginseng is one kind of the most precious herbal medicine in China. A variety of leaf diseases easily occur and disperse widely during the planting process at present. However, there is no universal detection of leaf diseases so far, due to the diversity and complexity of the leaf diseases of Panax notoginseng. Alternatively, a target detection technology using deep learning has been applied for detection tasks in various fields. This technology can be widely expected to greatly improve the efficiency of detection tasks for smart agriculture. In this study, an improved YOLOv3 (You Only Look Once v3) object detection (AD-YOLOv3) was proposed to locate and identify the complex diseases of Panax notoginseng leaves in dense and small areas. The attention feature pyramid in the AD-YOLOv3 was selected to replace the original in YOLOv3. Specifically, the channel attention module was used to compress each channel of the original feature map into a single value, then input into the fully connected layer and activate with the activation function for the weight of each feature channel, and finally the weight to perform the feature channel on the original feature map reorganization. As such, the important features were focused in the channel attention, further to discard the irrelevant for the less redundant interference in the multi-scale feature map during feature fusion. A double bottleneck layer was designed with two stacked blocks, where the first bottleneck block was added with the residual connection, and the second bottleneck block was the traditional structure in the attention feature pyramid. The resulting dual bottleneck block presented a higher performance than the raw. The dual bottleneck layer was further used to filter the features from the attention feature pyramid, indicating the better specificity of features and the robustness of detection. The attention feature pyramid and the double bottleneck layer were integrated to significantly improve the four performance indicators of the detection. Specifically, the AD-YOLOv3 had improved the overall accuracy, F1 accuracy, and mean average precision by 2.83, 1.68, and 1.47 percentage points, respectively, compared with YOLOv3. At the same time, the AD-YOLOv3 had improved the detection capabilities for each type of disease. The average precision of plague, rust disease, anthrax disease, powdery mildew, round spot disease, virosis, and all kinds of diseases were 78.0%, 79.5%, 86.7%, 69.0%, 85.9%, 84.2%, and 80.6%, respectively. The detection ability was significantly enhanced in the small and dense areas, as well as the anti-interference ability under complicated environments, such as fog, rain, and dark light. Correspondingly, there was a tradeoff between detection speed and accuracy in the AD-YOLOv3 with a simple structure, compared with the second-order object detection of Mask-RCNN. The modified AD-YOLOv3 can also be deployed to the server or client in the cloud detection of diseases in real time. The finding can provide a better intelligent detection for the leaf disease of Panax notoginseng.