Research on Common Tree Species Recognition by Faster R-CNN Based on Whole Tree Image

• Published in: 2021 IEEE 6th International Conference on Signal and Image Processing (ICSIP) pp. 28 - 32
• Main Authors: Li, Yu, Tang, Bo, Li, Jinda, Sun, Wei, Lin, Zhongkang, Luo, Qiang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.10.2021
• Subjects: deep learning; Faster R-CNN; Feature extraction; feature pyramid networks; Image recognition; Object detection; Real-time systems; Target recognition; Training; tree species recognition; Vegetation
• DOI: 10.1109/ICSIP52628.2021.9688693

In order to solve the problem of whole tree image recognition and detection, this paper introduces the object detection algorithm into the field of tree species recognition, and studies the Faster R-CNN tree species recognition method based on the whole tree image. Images of 10 types of common trees were collected under natural conditions in different weather conditions and made into a dataset containing 1000 pictures. The anchor frame size of the RPN network is adjusted according to the characteristics of the detection target in the dataset image, and the comparative experiments of the three feature extraction networks of VGG16, MobileNet-V2 and ResNet-50 are carried out. It is determined that ResNet-50 is the optimal feature extraction network. ResNet-50 is used as a model for feature extraction network training to recognize and detect 150 new images, and the detection accuracy reached 98%. The detection results show that the method can effectively detect the whole tree image, which can provide a reference for the application of object detection algorithm in the detection of the whole tree image.