Online Recognition and Counting with Deep Learning for Mango Images
Welcome to Chinese Journal of Tropical Crops,

Chinese Journal of Tropical Crops ›› 2020, Vol. 41 ›› Issue (3): 425-432.DOI: 10.3969/j.issn.1000-2561.2020.03.002

• Crop Culture and Nutrition, Genetic Breeding • Previous Articles     Next Articles

Online Recognition and Counting with Deep Learning for Mango Images

CEN Guanjun1,HUA Junda1,PAN Yiying1,LIU Dahe1,SU Beibei1,ZHONG Zheng1,ZHANG Liankuan1,*(),GAO Yan2,*()   

  1. 1. College of Mathematics and Informatics, South China Agricultural University, Guangzhou, Guangdong 510642, China
    2. Plant Protection Research Institute, Guangdong Academy of Agricultural Sciences / Guangdong Provincial Key Laboratory of High Technology for Plant Protection, Guangzhou, Guangdong 510640, China
  • Received:2019-03-14 Revised:2019-08-28 Online:2020-03-25 Published:2020-04-16
  • Contact: ZHANG Liankuan,GAO Yan

Abstract:

In order to realize the intelligent assessment of fruit yield, this paper carried out an research on the fruit recognition of Shengxin mango image in natural environment, and proposed an online method of recognition and counting with deep learning of mango images. Firstly, the recognition algorithm for mango image was realized using the Faster R-CNN Model, a deep learning framework. Secondly, an upload module for mango images based on Wechat applet and Web platform was developed, which can upload images to a server at all times and places. Thirdly, a server-client communication mode based on the TCP protocol and the Faster R-CNN assembly in MATLAB were adopted to construct an online analysis module, which realized the real-time and online recognition and counting for mango images. Finally, the recognition and counting results for a single picture, or all pictures in an orchard, were feedback to users through the Wechat applet program and Web page program, including the classification statistics of green mango and red mango. A total of 125 mango images in natural environment had been collected and analyzed using the method proposed in the paper. The results showed that the correct recognition and counting rate of mango fruit was 82.3%, among which the rate of missed detection and error detection was 11.7% and 8.6%, respectively, the average counting error was 4.2 and the average counting error rate was 7.9%. The experimental results demonstrated that, the method proposed in the paper was able to provide a scientific decision-making basis for the wisdom management of orchards through results quantity analysis obtained by images recognition and counting.

Key words: mango image, recognition and counting, Faster R-CNN, on-line method

CLC Number: