Enhancing YOLOv7 for Plant Organs Detection Using Attention-Gate Mechanism

Herbarium scans are valuable raw data for studying how plants adapt to climate change and respond to various factors. Characterization of plant traits from these images is important for investigating such questions, thereby supporting plant taxonomy and biodiversity description. However, processing these images for meaningful data extraction is challenging due to scale variance, complex backgrounds that contain annotations, and the variability in specimen color, shape, and orientation of specimens. In addition, the plant organs often appear compressed, deformed, or damaged, with overlapping occurrences that are common in scans. Traditional organ recognition techniques, while adept at segmenting discernible plant characteristics, are limited in herbarium scanning applications. Two automated methods for plant organ identification have been previously reported. However, they show limited effectiveness, especially for small organs. In this study we introduce YOLOv7-ag model, which is a novel model based on the YOLOv7 that incorporates an attention-gate mechanism, which enhances the detection of plant organs, including stems, leaves, flowers, fruits, and seeds. YOLOv7-ag significantly outperforms previous state of the art as well as the original YOLOv7 and YOLOv8 models with a precision and recall rate of 99.2% and 98.0%, respectively, particularly in identifying small plant organs.