Detection and Sizing of Durian using Zero-Shot Deep Learning Models

Abstract

Since 2017, up to 41% of Malaysia's land has been cultivated for durian, making it the most widely planted crop. The rapid increase in demand urges the authorities to search for a more systematic way to control durian cultivation and manage the productivity and quality of the fruit. This research paper proposes a deep-learning approach for detecting and sizing durian fruit in any given image. The aim is to develop zero-shot learning models that can accurately identify and measure the size of durian fruits in images, regardless of the image’s background. The proposed methodology leverages two cutting-edge models: Grounding DINO and Segment Anything (SAM), which are trained using a limited number of samples to learn the essential features of the fruit. The dataset used for training and testing the model includes various images of durian fruits captured from different sources. The effectiveness of the proposed model is evaluated by comparing it with the Segmentation Generative Pre-trained Transformers (SegGPT) model. The results show that the Grounding DINO model, which has a 92.5% detection accuracy, outperforms the SegGPT in terms of accuracy and efficiency. This research has significant implications for computer vision and agriculture, as it can facilitate automated detection and sizing of durian fruits, leading to improved yield estimation, quality control, and overall productivity.

Keywords

• durian
• grounding dino
• segment anything
• seggpt
• zero-shot learning