| dc.description.abstract | Climate change and population growth will put additional pressure on food security worldwide in the coming years. One way of combating this problem is to research and develop innovative seeds that could improve the yield. A crucial component of this is plant phenotyping - a scientific field aiming to characterize and quantify plants' physiological and physical features. This process can be automatized by employing computer vision algorithms. The focus of this work is to develop a computer vision pipeline capable of obtaining fruit's physical characteristics necessary for the task of plant phenotyping for the use case of zucchini fruits.
In this thesis, we developed, integrated, and evaluated a general segmentation pipeline that can be used for zucchini fruit segmentation and characterization. We improved the pipeline by adding methods to optimize and enhance the training process. That allowed us to obtain a model that can segment zucchini bodies and peduncles with a weighted F1 score of 0.85. Moreover, we expanded the zucchini fruit characterization step by developing an original fruit size estimation algorithm. This algorithm showed an average absolute error of 0.61 centimeters for zucchini length and 0.3 centimeters for zucchini diameters.
This thesis also addressed the problem of color reference target detection. Industry images containing CRTs are usually captured in the field in non-controlled conditions using devices like: DSLRs, cell phone cameras, drone cameras, and many more. Therefore, we constructed a novel CRT detection 2-stage method capable of extracting CRTs in a complex non-controlled environment which outputs data that can later be used for obtaining more accurate color information. | |