An edge-guided method to fruit segmentation in complex environments

Abstract

Accurate detection and segmentation of fruit is a key factor in the development of smart farming. Problems such as light variation, fruit overlap and leaf shading create a complex environment in orchards and have a significant impact on the development of smart farming. Many current deep learning-based segmentation methods do not make full use of edge information, resulting in inadequate sharpening of the fruit edges obtained from segmentation. To address this problem, an edge-guided based fruit segmentation method (EdgeSegNet) in complex environments is proposed by us. The method first performs feature extraction through the ResNet model as the backbone network, then integrates and refines the high-level semantic and spatial information through the Global Localization Module (GLM) and localizes potential targets in the target region with the help of the proposed Multi-Scale Localization Block (MSLB). Then Boundary Aware Module (BAM) sharpen the edges of potential targets by integrating the feature information of high and low layers, and finally get the accurate segmented image. The principle of the model is blurred positioning, precise sharpening, edge guiding. The experimental results showed that the method achieved an average MIoU of 0.909 and 0.942 on the apple and peach datasets of three different sizes, large, medium and small, respectively, outperforming several other state-of-the-art models in terms of accuracy and complexity as well as inference time.

Plant phenotyping relevance

果実を対象とした画像セグメンテーション手法を開発し、複数データセットで精度・複雑性・推論時間を比較評価しているため、植物表現型取得法が中心である。

Code and data availability

公開状態または取得可能な本文経路を確認できませんでした。

No evidence-backed public reproduction asset is currently recorded.