DetSSeg: A Selective On-Field Pomegranate Segmentation Approach

Segmenting pomegranate fruit on the field has many practical uses, including estimating yield, assessing fruit quality, identifying diseases, and optimizing harvest timing. This paper introduces a novel technique for detecting and segmenting pomegranate fruit from images. The proposed approach, named DetSSeg, combines YOLOv5s and UNet-MobileNetV2 models for selective on-field pomegranate fruit segmentation. This approach includes a unique training and inference flow that has significantly improved segmentation accuracy for on-the-field pomegranate images, which are often challenging due to diverse backgrounds, varying lighting conditions, and environmental factors. The proposed approach was evaluated through several case studies, and the results demonstrated a high dice coefficient of 98.18 % for outdoor pomegranate fruit segmentation, outperforming the single deep learning model that achieved a 99.32 % dice coefficient for indoor images but showing a lower accuracy of 87.56 % dice coefficient for on-field pomegranate fruit images. Moreover, the proposed approach outperformed other state-of-the-art segmentation models.