Berry Density Estimation With Deep Learning: estimating Density of Bilberry and Lingonberry Harvest with Object Detection
Pajula, Mikko (2022)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2022120927665
https://urn.fi/URN:NBN:fi:amk-2022120927665
Tiivistelmä
Currently, yield estimation of berry harvest is done manually by calculating berries in a predefined area. Manual counting is labour intensive and thus, the data created lacks scope. This thesis studied the method to estimate berry harvest as berry density value only with a consumer-level mobile phone camera. In addition, the thesis studied if berries detected from a top vegetation image could correlate with the berry density in the area defined by field measurement.
The density estimation process was twofold. Part one was to study the possibility of detecting lingonberry and bilberry from an image in the growth phases: flower, raw and ripe. YOLOv5 object detection was applied for this process. The original image was split into smaller images for deep learning to avert losing pixel information by scaling. Two split methods were tested.
Part two was to define the area for the image. Multiple methods were compared to define the most promising approach to estimating the area for berry density estimation. Field measurement values were used to describe the baseline area and to study the feasibility of determining berry density only from visible berries in the image.
The complete practical process of training the YOLOv5 object detection model was described in the thesis, from annotating datasets through editing and balancing datasets to training. Inadequacies in the dataset were described, and future improvements were suggested.
The density estimation process was twofold. Part one was to study the possibility of detecting lingonberry and bilberry from an image in the growth phases: flower, raw and ripe. YOLOv5 object detection was applied for this process. The original image was split into smaller images for deep learning to avert losing pixel information by scaling. Two split methods were tested.
Part two was to define the area for the image. Multiple methods were compared to define the most promising approach to estimating the area for berry density estimation. Field measurement values were used to describe the baseline area and to study the feasibility of determining berry density only from visible berries in the image.
The complete practical process of training the YOLOv5 object detection model was described in the thesis, from annotating datasets through editing and balancing datasets to training. Inadequacies in the dataset were described, and future improvements were suggested.