Semantic Segmentation of Germinated Oil Palm Seeds Based on Deep Convolutional Neural Networks with a Novel Channel Attention Mechanism

Abstract

Oil palm is a high value crop with an estimated 5% yearly replanting rate. To dispatch high quality seeds, stringent culling upon the germinated seeds is necessary. The shape of germinated part of an oil palm seed is a key manual criterion for distinguishing good seeds from bad ones. Accurate segmentation of the germinated part would serve as an important preprocessing step for automatic phenotypic analysis and quality classification of germinated oil palm seeds. In this paper, we pioneer the study of semantic segmentation of germinated oil palm seeds by convolutional neural networks (CNNs). Leveraging the state-of-the-art ‘SE-ResNext + U-Net’ architecture for image segmentation, we propose two modifications to address the difficulty of accurately segmenting the germinated part of a seed since it is much smaller compared to the seed body. Firstly, we design local spatial channel attention (LS-SE) to replace the Squeez-Excitation (SE) module to retain local information of a feature channel. Then we pass the features generated by the encoder to the same level of the decoder part twice along the decoding direction (DC-UNet) to retain the original features. This helped address the problem where the edge segmentation details of oil palm seeds require higher-resolution detail information to improve the segmentation accuracy. In addition, the number of parameters of the proposed DC-UNet is much smaller than that of other state-of-the art U-Net variants such as Unet++, significantly reducing the training time. Our proposed DC-UNet with LS-SE obtained an MIOU that is 1.2% higher than U-Net, with a clearly better visual segmentation around the boundaries of the germinated parts.

Plant phenotyping relevance

発芽種子の発芽部位という植物器官形質を画像から抽出するセグメンテーション手法を開発・評価しており、フェノタイピング解析の前処理手法が中心である。

Code and data availability

The paper uses a private, manually annotated oil palm seed image dataset (1752 training / 401 test images) and PyTorch-implemented models (DC-UNet, LS-SE), but no blocks contain any public data deposit, code repository, or availability statement. The authors explicitly describe the dataset as private, and no authors'-p

No evidence-backed public reproduction asset is currently recorded.