BiSeNeXt: a yam leaf and disease segmentation method based on an improved BiSeNetV2 in complex scenes.

Abstract

Introduction Yam is an important medicinal and edible crop, but its quality and yield are greatly affected by leaf diseases. Currently, research on yam leaf disease segmentation remains unexplored. Challenges like leaf overlapping, uneven lighting and irregular disease spots in complex environments limit segmentation accuracy. Methods To address these challenges, this paper introduces the first yam leaf disease segmentation dataset and proposes BiSeNeXt, an enhanced method based on BiSeNetV2. Firstly, dynamic feature extraction block (DFEB) enhances the precision of leaf and disease edge pixels and reduces lesion omission through dynamic receptive-field convolution (DRFConv) and pixel shuffle (PixelShuffle) downsampling. Secondly, efficient asymmetric multi-scale attention (EAMA) effectively alleviates the problem of lesion adhesion by combining asymmetric convolution with a multi-scale parallel structure. Finally, PointRefine decoder adaptively selects uncertain points in the image predictions and refines them point-by-point, producing accurate segmentation of leaves and spots. Results Experimental results indicated that the approach achieved a 97.04% intersection over union (IoU) for leaf segmentation and an 84.75% IoU for disease segmentation. Compared to DeepLabV3+, the proposed method improves the IoU of leaf and disease segmentation by 2.22% and 5.58%, respectively. Additionally, the FLOPs and total number of parameters of the proposed method require only 11.81% and 7.81% of DeepLabV3+, respectively. Discussion Therefore, the proposed method can efficiently and accurately extract yam leaf spots in complex scenes, providing a solid foundation for analyzing yam leaves and diseases.

Plant phenotyping relevance

ヤム葉と病斑を画像から分割・抽出する手法とデータセットを開発し、性能比較まで行っており、植物の病害状態を取得する方法が中心である。

Code and data availability

The paper's authors publicly released their self-constructed yam leaf disease segmentation dataset (1,097 annotated images of anthracnose, brown spot, and gray spot) via a Google Drive link in the Data availability statement. No code or trained model deposit is explicitly stated.