Texture feature guided attention based fusion representations for crop leaf disease detection

Abstract

The goal of automated crop leaf disease detection (CLDD) is to extract fine-grain visual features from images for classification. Vision transformers (ViTs) and attention-based models have improved classification accuracies by narrowing down the receptive fields of visual features, though often at the expense of computations and robustness. Generalizing visual features becomes difficult due to lacking diversification and overrepresentation of healthy leaves over diseased ones, causing overfitting and inherently limiting ViT’s capacity. Vision Transformers (ViTs) typically rely on single feature projection onto query, key, and value embeddings for self-attention calculations. In contrast, we introduce multi-feature projection. The texture representations extracted from RGB images are fed into key and value components, while the query takes RGB-coded features. Attention is the SoftMax-activated dot product computed on the three components. Simultaneously, a residual branch integrates RGB features with the attention vectors, producing moderately robust and interpretable compositions referred to as Texture Guided Visual Attention (TGVA) features. These TGVA features are integrated back into the backbone classifier network. Evaluating the Texture Guided Visual Attention neural network (TGVAnn) on the most challenging plant leaf datasets, sugarcane and maize, demonstrates its superior performance, achieving a 9% improvement in classification accuracy over traditional ViT models. Furthermore, TGVAnn shows a 20%–70% reduction in theoretical computational requirements (GFLOPs) relative to comparable baselines under a common reporting convention, supporting efficiency and scalability. This method outperforms similar approaches in both accuracy and robustness. In conclusion, the TGVA features derived from the multi-feature attention module effectively enhance the robustness of backbone image classifiers. This improvement is achieved with only a minimal increase in computational overhead of 13 M-parameters, making the approach both efficient and suitable for edge deployment scenarios.

Plant phenotyping relevance

植物葉画像から病害状態を推定する画像解析手法を開発し、複数データセットで性能評価しており、フェノタイピング手法が中心である。

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.