KDOSS-net：作物と雑草の画像のための知識蒸留ベースのアウトペインティングおよびセマンティックセグメンテーションネットワーク

Weed management plays a crucial role in increasing crop yields. Semantic segmentation, which classifies each pixel in an image captured by a camera into categories such as crops, weeds, and background, is a widely used method in this context. However, conventional semantic segmentation methods rely solely on pixel information within the camera's field of view (FOV), hindering their ability to detect weeds outside the visible area. This limitation can lead to incomplete weed removal and inefficient herbicide application. Incorporating information beyond the FOV in crop and weed segmentation is therefore essential for effective herbicide usage. Nevertheless, existing research on crop and weed segmentation has largely overlooked this limitation. To address this issue, we propose the knowledge distillation-based outpainting and semantic segmentation network (KDOSS-Net) for crop and weed images, a novel framework that enhances segmentation accuracy by leveraging information beyond the FOV. KDOSS-Net consists of two parts: the object prediction-guided outpainting and semantic segmentation network (OPOSS-Net), which serves as the teacher model by restoring areas outside the FOV and performing semantic segmentation, and the semantic segmentation without outpainting network (SSWO-Net), which serves as the student model, directly performing segmentation without outpainting. Through knowledge distillation (KD), the student model learns from the teacher's outputs, which results in a lightweight yet highly accurate segmentation network that is suitable for deployment on agricultural robots with limited computing power. Experiments on three public datasets-Rice seedling and weed, CWFID, and BoniRob-yielded mean intersection over union (mIOU) scores of 0.6315, 0.7101, and 0.7524, respectively. These results demonstrate that KDOSS-Net achieves higher accuracy than existing state-of-the-art (SOTA) segmentation models while significantly reducing computational overhead. Furthermore, the weed information extracted using our method is automatically linked as input to the open-source large language and vision assistant (LLaVA), enabling the development of a system that recommends optimal herbicide strategies tailored to the detected weed class.