Improved YOLOv8-Based Algorithm for Citrus Leaf Disease Detection

In view of the small difference between citrus leaf diseases which can lead to false inspection and missed inspection, an improved YOLOv8 citrus leaf disease detection algorithm is proposed. The proposed approach uses YOLOv8n as the base model and introduces adaptive convolution into the Backbone, allowing the model to dynamically prioritize different disease features. This modification increases the model’s parameter count without adding to the computational cost in terms of floating-point operations. Next, a fast convolution layer is implemented to replace the original C2f module, improving both detection accuracy and computational efficiency. In addition, a multi-scale fusion attention module is incorporated into the Neck section, which effectively integrates disease-related information from different receptive fields, thereby boosting the model’s ability to detect a wider variety of diseases. To further enhance adaptability, the traditional Intersection over Union (IoU) is replaced with the Wise-IoU loss function, improving the model’s performance across different disease types. The algorithm model uses a self-built data set containing 1,810 images, with 481 images of Huanglongbing, 430 images of magnesium deficiency, 451 images of anthracnose, and 448 images of black spot. In order to maintain the randomness of the experimental data, this experiment divided the data set into training set, validation set and test set with a ratio of 7:2:1, which contained 1267 images, 362 images and 181 images respectively. Experimental results show that the improved algorithm achieves an accuracy of 93.2% on the test set, representing a 1.2 percentage increase over the original model, while also reducing the parameter volume and computational complexity by 7% and 7.4% respectively. The improved model notably decreases the missed detection rate under occlusion and improves the detection of small targets. Furthermore, it strikes an effective balance between enhancing detection accuracy, reducing model complexity, and maintaining a lightweight architecture, making them well-suited for efficient citrus leaf disease detection.