Regresar

Mask R-CNN and YOLOv8 Comparison to Perform Tomato Maturity Recognition Task

Abstract:

This work explores the segmentation and detection of tomatoes in different maturity states for harvesting prediction by using the laboro tomato dataset to train a mask R-CNN and a YOLOv8 architecture. This work aims to test the mask R-CNN architecture and the proposed methodology efficiency on the benchmark paper [12]. The evaluation metric intersection over union (IoU) 0.5 showed an average precision of 67.2% with a recall of 78.9% over the laboro tomato dataset and an IoU average precision of 92.1% with a recall of 91.4% over the same dataset. The benchmark paper authors perform segmentation and classification in a separate process using color analysis algorithms and use the determination coefficient (R<sup>2</sup> ) for how accurately the tomato was set into the three maturity classes. The results show that the state-of-the-art YOLOv8 has a R<sup>2</sup> of 0.809, 0.897, and 0.968 in the ripe, half-ripe, and green categories, respectively. However, the Mask R-CNN results are acceptable, with 0.819, 0.809, and 0.893 in the ripe, half-ripe, and green categories, respectively. The YOLOv8 model performed better than the one used in the benchmark paper by detecting, segmenting, and classifying tomatoes. Moreover, the color-analysis technique used in the benchmark paper results inefficiently because the classification results showed no linear relation between the predictions and the real values.