Litchi Anthracnose Dataset

This dataset was collected from the core litchi-producing areas in Maoming City, Guangdong Province, China, during the litchi ripening seasons (May-June) of 2024 and 2025. The dataset focuses on anthracnose disease in two market-dominant litchi varieties, namely "Baitangying" and "Feizixiao". Images were captured from different angles targeting three levels of anthracnose: Mild, Moderate, and Severe.The collection process was based on natural indoor light, and the background was not completely uniform and simple, but it truly reflects the actual state of litchi affected by anthracnose. To truly reflect the impact of device diversity in practical applications, four mainstream smartphones, including iPhone 12, Honor 50, Honor X50, and realme GT Neo (Speed Edition), were used for image collection. All devices uniformly used the main camera (with a focal length of approximately 26mm), and the core shooting parameters were set to automatic mode (ISO automatically adjusts ambient light sensitivity, white balance (AWB) automatically calibrates colors, and exposure mode (P/Auto) automatically controls brightness). The captured images were directly stored as high-quality JPEG files at the device's native default resolution.The original dataset contains 3644 visible light raw images of litchi fruit anthracnose. To improve the model's generalization ability and robustness to complex and variable lighting conditions, three data enhancement strategies (front light, backlight, and side light) were used to expand the original dataset, and the total size of the final dataset was expanded to 14576 images. The expanded dataset was divided into a training set (10203 images), a validation set (2915 images), and a test set (1458 images) in the proportion of 70%, 20%, and 10%.All images were professionally annotated using the Labelimg tool to generate xml files containing target bounding box coordinates and category labels. The annotation process followed unified visual target annotation specifications, and three-level quality inspection (self-inspection by annotators and review by team leaders) was conducted to ensure annotation consistency. Meanwhile, an automated conversion tool was developed to convert xml annotations into YOLO-format txt files, providing professional and reliable data support for the training and validation of litchi disease grading detection models.