Method for identifying chili pepper leaf diseases and pests in natural environments based on an improved YOLOv8n

Chili pepper leaf diseases and insect pests have seriously threatened the yield and quality under natural cultivation. While accurate identification is often confined to complex backgrounds, variable illumination, and small or partially occluded targets. In this study, an improved YOLOv8n was proposed for the recognition accuracy of chili pepper leaf diseases and insect pests in natural scenes. The images were collected from major chili pepper-producing areas under different field conditions, including sunny and cloudy weather, front-light and backlight environments, and different shooting distances. The original dataset contained 1 283 annotated images covering four categories, namely viral disease, epidemic disease, thrips, and green worms. A series of processing was conducted to avoid data leakage. The original images were first divided into training, validation, and test sets at a ratio of 8:1:1, and data augmentation was then applied only to the training set. Data augmentation included random flipping, random affine transformation, scaling, translation, and image blurring, which improved the diversity of training samples and the adaptability of the model to natural field conditions. Three improvements were introduced into the original YOLOv8n model. First, a Selective Kernel (SK) module was embedded into the backbone network to strengthen the feature representation for targets at different scales, where the convolution branches were adaptively selected with different receptive fields. Second, part of the C2f structure in the backbone was replaced with a Biformer module. The key information regions were obtained to suppress redundant background interference using a bi-level routing attention mechanism, thereby enhancing the discrimination of disease and pest targets in complex scenes. Third, the Complete Intersection over Union (CIoU) loss was replaced with the Focal-Inner Complete Intersection over Union (FI-CIoU) loss. The contribution of difficult samples was improved during bounding box regression. The localization performance was also enhanced for small and partially occluded targets. Experimental results showed that the improved model achieved better performance than the original YOLOv8n model on the dataset. Precision increased from 63.8% to 66.9%, recall increased from 58.1% to 59.9%, and mean average precision increased from 59.7% to 65.1%, where it improved by 3.1, 1.8, and 5.4 percentage points, respectively. Visual detection indicated that the improved model effectively reduced missed and false detections in complex natural scenes, indicating better adaptability to dense leaf backgrounds, small targets, and partially occluded targets. Overall, the better performance was obtained in recognizing the chili pepper leaf diseases and insect pests under natural conditions. The findings can also provide a strong technical reference to monitor chili pepper diseases and insect pests.