Improved YOLOv8 Algorithm was Used to Segment Cucumber Seedlings Under Complex Artificial Light Conditions

With the widespread application and rapid proliferation of agricultural solutions in plant factories, "intelligent environmental control technology" has become an important research topic as one of the core functions of plant factories. Accurately obtaining the growth status of plants is the core of intelligent control technology and has been capturing more and more researchers’ attention. In the study of detecting the growth status of cucumber plants, the accurate segmentation of cucumber leaves is crucial. However, under fully artificial lighting conditions in plant factories, the complexity of light is the main challenge for leaf segmentation. Aiming at the challenging problem of cucumber seedling leaf segmentation under a complex artificial lighting environment,this study proposes an improved complex lighting YOLOv8 (CL-YOLOv8) model based on YOLOv8. First, the C2f module in the YOLOv8 backbone network is optimized by replacing the bottleneck with the multi-scale feature module (MSFM) we designed. The convolution kernels of different sizes are employed to capture features at multiple scales in the image, thereby processing target and background information more effectively. Simultaneously, the Efficient Multi-Scale Attention (EMA) mechanism is introduced after MSFM, enabling the model to flexibly adjust the attention allocated to different regions within the image, thus enhancing its robustness to complex lighting variations. Finally, the efficient sub-pixel upsampling block (ESPUB) is designed to improve image resolution without compromising computational efficiency, thereby enhancing the model’s adaptability. The verification results demonstrate that the improved CL-YOLOv8 model, based on the YOLOv8 network, enhances the mAP@50 and mAP@50-95 metrics by 2.412% and 4.148%, respectively, achieving real-time and accurate segmentation of cucumber leaves. The effectiveness and superiority of this approach are validated through comparisons with other widely used segmentation detection algorithms.