MobUNet: Utilizing Deep Learning for Segmenting Cucumber Leaves

Abstract

Plant image segmentation is challenging due to overlapping leaves and complex image backgrounds. Consequently, the segmentation model has some challenges recognizing the leaves, further affecting the segmentation performance. This study proposes a deep learning method called MobUNet, using U-Net, and employs MobileNetV2 as an encoder to overcome the problems for cucumber leaf segmentation. Around 145 leaf images with complex backgrounds are collected at the cucumber farm and annotated for ground truth data. The experiment uses the ratio of 80:20 for training and testing sets, and some hyperparameters are modified to achieve a good segmentation result. The segmentation results are subject to several metrics: accuracy, Dice score, IoU, Dice loss, Jaccard distance, and Hausdorff distance. The experimental results for segmentation accuracy, Dice score, and IoU were 93.23%, 91.30%, and 85.03%, respectively. An analysis was conducted to create a benchmark in segmentation performance, utilizing the U-Net baseline, MobileNetV1, and MobileNetV2, which use the same dataset. Despite the complex background, MobUNet can successfully segment the cucumber leaf images compared to the other models. The MobUNet showed the closest Hausdorff distance value to the origin point, measuring at 0.0001; hence, it demonstrates high quality and accuracy in the segmentation.