Apple, peach, and pear flower detection using semantic segmentation network and shape constraint level set

Abstract

In fruit production, the number of flowers plays a critical factor in crop management decision in an orchard. This paper proposes an automated apple, peach and pear flower detection method under varied environments. The semantic segmentation network DeepLab-ResNet is fine-tuned using apple flower dataset and used in detection for apple, peach and pear flower datasets. On the assumption that the network can roughly locate the flower object and there is distinct color difference between the flower and the surrounding background, an active contour model is used to refine the coarse segmentation results of the network. Specifically, the result from the network presents a shape constraint in the active contour model. The method is tested on four public available image datasets of apple, peach and pear flowers under different environments. The experimental results reveal that the level set model can improve the segmentation result of the semantic segmentation network, especially when the network is generalized to datasets other than those used in network training. Our method achieves a F₁ score at pixel-level up to 89.6% on one of the apple dataset and an average F₁ score of 80.9% on the peach, pear and another apple datasets, which are 6% and 5% higher than the previous state-of-the-art region growing refinement method on the same datasets, respectively.

Plant phenotyping relevance

果実生産管理に用いる花数という植物器官形質を、画像セグメンテーションで自動抽出する手法を開発・評価しており、フェノタイピング手法が中心です。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.