Click on “Download PDF” for the PDF version or on the title for the HTML version.


If you are not an ASABE member or if your employer has not arranged for access to the full-text, Click here for options.

Anchor-free deep convolutional neural network for plant and plant organ detection and counting

Published by the American Society of Agricultural and Biological Engineers, St. Joseph, Michigan www.asabe.org

Citation:  2021 ASABE Annual International Virtual Meeting  2100738.(doi:10.13031/aim.202100738)
Authors:   Chenjiao Tan, Changying Li, Dongjian He, Huaibo Song
Keywords:   Anchor free, counting, deep convolutional network, detection, plant and plant organ

Abstract. Accurately counting the number of plants and plant organs in natural environments is essential for breeders and growers. Plant count can provide the emergence rate for breeding a new cultivar and evaluating the necessity of replanting. Plant organ count, especially the flower count, reflects the crop yields of particular genotypes and helps growers know yields in advance. In order to avoid repeated counting in adjacent frames of videos collected in natural environments, we describe a deep convolutional neural network (CenterTrack)-based tracking method for cotton seedling and flower counting. The network is extended from our modified CenterNet, which is an anchor-free object detector. CenterTrack takes the current frame, the previous frame and detection results from previous frame as inputs and predicts the detections of the current frame and offsets of each detection. Displacements are used to associate the same seedling. Unmatched detections in previous frames are removed from the track list and unmatched detections in the current frame are added to the track list. Matched detections are saved with new positions in the existing track list. The modified CenterNet detector achieved high accuracy on both seedling and flower datasets. Overall, AP50 of the final model was 0.952. Experimental results showed that seedling and flower counts highly correlated with those of manual counts (=0.95) and the average error of all 100 testing videos was 9.26%. The anchor-free deep convolution neural network-based tracking method provides an automatic tracking and counting in video frames, which will significantly benefit plant breeding and crop management.

(Download PDF)    (Export to EndNotes)