An improved U-Net network-based quantitative analysis of melon fruit phenotypic characteristics

Melon fruit phenotype has rich genetic variation and is an important target trait for breeding and commerciality. Existing measurement techniques have problems such as low detection accuracy, low efficiency and vague quantification methods. The study uses a deep learning approach to build an improve...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:Journal of food measurement & characterization 2022-10, Vol.16 (5), p.4198-4207
Hauptverfasser: Qian, Chunyang, Liu, Haolin, Du, Taihang, Sun, Shuguang, Liu, Wei, Zhang, Ruowei
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Beschreibung
Zusammenfassung:Melon fruit phenotype has rich genetic variation and is an important target trait for breeding and commerciality. Existing measurement techniques have problems such as low detection accuracy, low efficiency and vague quantification methods. The study uses a deep learning approach to build an improved U-Net network for accurate segmentation of melon pericarp and seed cavity characteristics, as well as quantitative computation of phenotypic characteristics. The improved network uses VGG16 to replace the encoder in the original U-Net network, migrates the pre-training weights under the PASCAL VOC dataset, and adds a Repeated Criss-Cross Attention to the U-Net network skip connection to improve the network training efficiency and enhance the acquisition of image contextual information to solve the problem of few melon samples and poor fruit skin cavity segmentation. The segmentation results were further quantified to obtain precise parameters of melon fruit phenotypic characteristics. The experiments showed that the segmentation accuracy indexes of the improved U-Net network were 98.83% for Mean Pixel Accuracy and 96.67% for Mean Intersection over Union, which were higher than those of other comparative algorithms. The relative error of each quantitative measurement parameter is less than 5%, and the repeat accuracy is less than 0.1%, which has high accuracy and robustness. It meets the requirements of rapid detection of melon fruit characteristics and provides a powerful help for the accurate acquisition of its germplasm resource information.
ISSN:2193-4126
2193-4134
DOI:10.1007/s11694-022-01519-7