Developing a high-throughput phenotyping software for eggplant through trichome analysis using deep learning
2024
Doria, A.B.
Computer vision is widely used in precision agriculture to lessen the human tasks and errors brought about by human limitations. One of the tedious parts of agriculture experiments in phenotyping tasks specifically in trichome counting which is used for developing eggplant varieties resistant to pests. To address this problem, an Eggplant Trichome Detection model is developed using the YOLOv4 tool to automate the detection and counting of trichomes in microscopic images. Applications for labeling and training were also created to help researcher develop their own detection model based on the availability of the images. Different images datasets with specific groups of trichomes were created to experiment with various image-pre-processing techniques namely the Contrast Limited Adaptive Histogram Equalization (CLAHE) and Gamma correction to come up with a better trichome detection model. Using the Mean Average Precision (MaP) calculation in measuring the performance of the detection, gamma-corrected images got the highest precision at 85.71% at 50% intersection over union (IoU) threshold while the non-pre-processed images had the highest precision at 92.19% on 30% IoU. On the other hand, CLACHE got the lowest precision at 67.52% and 72.59% under the 50% and 30% IoU threshold respectively. On the other hand, an average of 0.18 seconds is allotted to analyze a certain image using the built software as compared to 30 to 120 seconds for manual counting and detection.
显示更多 [+] 显示较少 [-]