关键词: K-means, Des Faster-RCNN, 迁移学习, 单木检测, 锚点

Abstract:

To study the single tree detection technology in the process of intelligent industrialization and precise management of fruit forests, this paper collected high spatial resolution remote sensing data of navel orange orchards in southern Jiangxi Province, and used K-means clustering analysis anchor (anchor points) to improve network parameters; proposed an improved network structure Des Faster-RCNN network to ensure the multiplexing and fusion of shallow features during the process of network convolution extracted features; added transfer learning in the training process to achieve fine-tuning. The robustness of detected small targets and densely adhered targets was enhanced. The experimental results showed that the detection accuracy of the Des Faster-RCNN network was improved by adding K-means on the test set was 97.68%, the recall rate was 97.99%, and the F₁ value was 0.9783, increased by 8.01%, 8.14% and 0.081, respectively, compared with the those of Faster-RCNN network. And compared with other detection frameworks and traditional fruit forest detection research methods, the detection accuracy was improved by about 10%. The average detection time of a single image was about 0.138 s. In conclusion, the Des Faster-RCNN network improved by adding K-means proposed in this paper can adapt to the detection task of small and dense targets in complex environments, and obtain higher detection accuracy and robustness.

Key words: K-means, Des Faster-RCNN, transfer learning, single tree recognition, anchor