摘要:無人航空機(UAV)を用いて撮影したトドマツの樹冠から,畳み込みニューラルネットワークを基にした画像認識アルゴリズムであるYou Only Look Once(YOLO)v4を用いて球果を検出するモデルを構築し,その精度を検証した。356枚,合計6,138個の球果が写った画像で学習を行い,構築したモデルを92枚,合計1,692個の球果が写った検証用画像に適用した結果,88.5%のaverage precision(AP)が得られた。一方で,白く円形の小型物体を誤検出したfalse positiveや,密集した球果を検出できなかったfalse negativeの事例があり,これらの解決は今後の課題と考えられた。YOLOv4を用いてUAV撮影画像からトドマツの球果を検出することは可能であり,今後,採種園で球果を生産した個体の確認に有用と期待される。
其他摘要:An object detection model for detecting the cones of Abies sachalinensis on the tree crown using images shot by unmanned aerial vehicles (UAV) was developed. We used the image recognition algorithm "You Only Look Once (YOLO) v4" based on a convolutional neural network and examined its accuracy. Training was performed using 356 pictures with 6,138 cones, and the constructed model was adapted to 92 validation pictures with 1,692 cones. As a result, an average precision (AP) of 88.5% was obtained. However, small white round objects were often detected as cones (false positives) and densely situated cones were not detected (false negative). Improvement of those misdetections will be future subject. We conclude that cone detection of A. sachalinensis using YOLOv4 is possible, and the model will be useful to confirm cone producing individuals in seed orchards.