Lightweight tomato disease recognition in complex environments based on improved YOLOv8

With the continuous expansion of tomato cultivation, early identification and accurate control of diseases are crucial for reducing economic losses and ensuring food security. However, in natural scenarios such as greenhouses and fields, challenges like complex backgrounds, drastic lighting changes, and limited computing power of edge devices make it difficult to achieve both high accuracy and real-time detection. To address these issues, this article proposes a lightweight real-time detection method based on an improved YOLOv8 and constructs a hybrid scenario dataset, TomatoDataset, for validation. The dataset includes samples from complex field environments and controlled settings to ensure the visibility of fine-grained characteristics of tiny disease spots. Methodologically, a lightweight backbone utilizing Inverted Residual Blocks and Depthwise Separable Convolutions is designed to significantly compress the parameter space while preserving feature expression capabilities. Furthermore, a Coordinate Attention (CA) mechanism is embedded in the feature fusion stage to enhance directional perception, and the Scale–Sensitive Intersection over Union (SIoU) loss function is introduced to optimize bounding box regression for tiny targets. Experimental results demonstrate that the proposed model achieves an F1-score of 74.65%, mean average precision (mAP)@0.5 of 79.5%, and mAP@0.5:0.95 of 41.6% on the TomatoDataset. On the NVIDIA Jetson Nano platform, the per-image inference latency is only 33.6 ms. Compared with the mainstream lightweight model YOLOv8n, our model reduces the parameter volume by 46.9% while improving the F1-score by 1.9%. This study achieves a coordinated optimization of detection accuracy and computational efficiency, providing a viable solution for real-time tomato disease monitoring on edge devices.