High-throughput phenotyping technology for corn ears

Phenotyping;Corn ears;High-throughput;Computer vision;Agriculture

Zhao, Wei ; Grift ; Tony E.

关键词: Phenotyping;    Corn ears;    High-throughput;    Computer vision;    Agriculture;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/72828/Wei_Zhao.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

The phenotype of any organism, or as in this case, plants, includes traits or characteristics that can be measured using a technical procedure. Phenotyping is an important activity in plant breeding, since it gives breeders an observable representation of the plant’s genetic code, which is called the genotype. The word phenotype originates from the Greek word “phainein” which means “to show” and the word “typos” which means “type”. Ideally, the development of phenotyping technologies should be in lockstep with genotyping technologies, but unfortunately it is not; currently there exists a major discrepancy between the technological sophistication of genotyping versus phenotyping, and the gap is getting wider. Whereas genotyping has become a high-throughput low-cost standardized procedure, phenotyping still comprises ample manual measurements which are time consuming, tedious, and error prone. The project as conducted here aims at alleviating this problem; To aid breeders, a method was devised that allows for high-throughput phenotyping of corn ears, based on an existing imaging arrangement that produces frontal views of the ears. This thesis describes the development of machine vision algorithms that measure overall ear parameters such as ear length, ear diameter, and cap percentage (the proportion of the ear that features kernels versus the barren area). The main image processing functions used here were segmentation, skewness correction, morphological operation and image registration. To obtain a kernel count, an “ear map” was constructed using both a morphological operation and a feature matching operation. The main challenge for the morphological operation was to accurately select only kernel rows that are frontally exposed in each single image. This issue is addressed in this project by developing an algorithm of shadow recognition. The main challenge for the feature-matching operation was to detect and match image feature points. This issue was addressed by applying the algorithms of Harris’s Conner detection and SIFT descriptor. Once the ear map is created, many other morphological kernel parameters (area, location, circumference, to name a few) can be determined. Remaining challenges in this research are pointed out, including sample choice, apparatus modification and algorithm improvement. Suggestions and recommendations for future work are also provided.