Authors

Abstract

The highly similar visual appearance among different cultivated plant species makes fine-grained plant cultivar retrieval a challenging task. Considerable efforts have been devoted to this problem, and significant progress has been achieved in recent decades. This paper proposes a simple and effective method for fine-grained plant cultivar retrieval. The main contributions are threefold. First, experimental analysis indicates that image resolution plays a crucial role in fine-grained plant retrieval, with 896×896 pixels representing the most cost-effective resolution. Second, a radial basis kernel function is employed to capture the nonlinear channel correlation of the feature map, enabling the extraction of more discriminative features. In addition, Log-TiedRank is applied to improve robustness to noise and to obtain a more compact representation. Finally, two types of deep features extracted from two convolutional neural networks are fused to further enhance retrieval performance. Compared with state-of-the-art methods, the proposed approach improves the retrieval rate by 15.71%, 15.95%, 14.02%, and 8% on the SoyCultivar200 dataset, 4.68% on PeanCultivar100, and 4.01% on the Mulberry dataset, demonstrating the effectiveness and superiority of the proposed method.

Abstract in Chinese

不同栽培品种植物在视觉外观上的高度相似性，使得精细植物栽培品种检索成为一项具有挑战性的任务。过去几十年中，研究者们投入了大量努力并取得了显著进展。本文提出了一种简洁有效的精细植物栽培品种检索方法，主要贡献体现在三个方面。首先，通过实验发现图像分辨率在精细植物检索中起着至关重要的作用，而896×896像素可能是最具成本效益的分辨率选择。其次，我们采用径向基核函数捕捉特征图的非线性通道关联信息，旨在提取更具判别力的特征。同时引入Log-TiedRank方法提升紧凑表征对噪声的鲁棒性。最后，通过融合两种卷积网络提取的深度特征进一步优化检索性能。实验结果表明，本方法在SoyCultivar200数据集上检索率较现有最优方法提升15.71%、15.95%、14.02%和8%，在PeanCultivar100和Mulberry数据集上分别提升了4.68%和4.01%，充分证明了该方法的优越性与有效性。