TasselNetV2++: A dual-branch network incorporating branch-level transfer learning and multilayer fusion for plant counting

Abstract

Plant counting plays an important role in evaluating planter effectiveness, assessing seed quality, devising agricultural management plans, and estimating crop yields. Given its significance and the ease of acquiring agricultural images, the development of an end-to-end image-based plant counting model applicable across diverse agricultural settings is crucial. The proposed TasselNetV2++, an improved version of TasselNetV2+ for plant counting, introduces notable enhancements to its encoder and counter while maintaining the existing normalizer. In the encoder, we designed a dual-branch architecture, with one branch being a customized YOLOv5s backbone and the other branch being the original encoder equipped with an attention mechanism. It is precisely the branch-level transfer learning, coupled with multilayer fusion, within the dual-branch architecture that significantly enhances the feature extraction capability of the network across a wide range of scenarios. Moreover, the counter has been enhanced with an attention mechanism that recalibrates its focus on crucial spatial locations and channel-wise features following average pooling. Experimental results demonstrate that TasselNetV2++ outperforms its predecessor across multiple counting tasks. Compared to TasselNetV2+, TasselNetV2++ achieves a substantial reduction in relative root mean squared error (rRMSE). Specifically, it brings a 33.3% relative decrease of rRMSE on the soybean seedlings counting dataset, 8.4% on the wheat ears detection dataset, 28.6% on the maize tassels counting dataset, and 18.0% on the sorghum heads counting dataset. Notably, ablation experiment demonstrates the indispensability of the branch-level transfer learning in achieving precise plant counting. Branch-level transfer learning achieves a notable relative decrease in rRMSE of 31.4% for soybean seedlings, 7.9% for wheat tassels, 36.5% for maize tassels, and 2.0% for sorghum heads. The proposed TasselNetV2++ attains remarkable advancements and introduces a straightforward yet highly effective branch-level transfer learning strategy.

Plant phenotyping relevance

植物個体・器官の画像ベース計数モデルを開発し、複数データセットで性能比較とアブレーション評価を行っており、表現型取得・抽出法が中心である。

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.