MHW-PD: A robust rice panicles counting algorithm based on deep learning and multi-scale hybrid window

Abstract In-field assessment of rice panicle yields accurately and automatically has been one of the key ways to realize high-throughput rice breeding in the modern smart farming. However, practical rice fields normally consist of many different, often very small sizes of panicles, particularly when large numbers of panicles are captured in the imagery. In these cases, the integrity of panicle feature is difficult to extract due to the limited panicle original information and substantial clutters caused by heavily compacted leaves and stems, which results in poor counting efficacy. In this paper, we propose a simple, yet effective method termed as Multi-Scale Hybrid Window Panicle Detect (MHW-PD), which focuses on enhance the panicle features to detect and count the large number of small-sized rice panicles in the in-field scene. On the basis of quantifying and analyzing the relationship among the receptive field, the size of input image and the average dimensions of panicles, the MHW-PD gives dynamic strategies for choosing the appropriate feature learning network and constructing adaptive multi-scale hybrid window (MHW), which maximizes the richness of panicle feature. Besides, a fusion algorithm is involved to remove the repeated counting of the broken panicles to get the final panicle number. With extensive experimental results, the MHW-PD has achieved ~87% of panicle counting accuracy; and the counting accuracy just decreases by ~8% when the number of panicles per image increases from 0 to 80, which shows better in stability than all the competing methods adopted in this work. The MHW-PD is demonstrated qualitatively and quantitatively that is able to deal with high density of panicles.