A Global Feature Fusion Network for Lettuce Growth Trait Detection

Lettuce growth traits are important biological attributes that directly reflect growth conditions. However, most existing approaches simply extract coarse features from RGB images and ignore the significantly varied appearance of different lettuce varieties at diverse growth phases, which brings about the loss of important information. To address these issues, we propose a novel lettuce growth-traits detection model, namely Global Feature Fusion Network (GFFN), based on dense connection and dilated convolution to fully utilize fine-grained and multi-level feature representations from RGB-D images. Firstly, RGB and depth images are combined through channel concatenation to provide rich, learnable information. Next, a dense extractor is proposed to perform progressively refined feature extraction, which gathers fine-grained local context from coarse lettuce representations. Then, a multi-scale receptor aims to merge multi-level feature representations and learn scale and location knowledge. Finally, extensive experiments show that GFFN achieves competitive performance compared to the other mainstream methods in detecting five primary attributes of lettuce growth traits.