Partial convolutional biformer: A transformer architecture for diagnosing crop diseases under complex backgrounds

Abstract

In agricultural scenarios, the images obtained are often affected by factors such as weather and environmental conditions, which can introduce varying levels of noise to the images. This requires computer vision models to possess a degree of robustness. Generally, model with strong robustness comes with higher computational complexity and model size, which place greater demands on hardware computing resources during deployment. Hence, this research proposes PConv BiFormer (PCBT) based on the BiFormer architecture for the purpose of identifying crop diseases. Prior to inputting images into the network, an additional convolutional layer is introduced to use feature maps generated through convolutional operations as the model’s input. Furthermore, the Depthwise Convolution in BiFormer is replaced with Partial Convolution (PConv) to encode relative positional information. The Convolutional Gated Linear Unit was introduced as the model’s channel mixer to filter global information, aiming to enhance the model’s robustness. PCBT-small has classification accuracies of 0.998, 0.878, and 0.919 on three datasets with computational load of 2.16G FLOPs and had a parameter count of 10.20M. PCBT maintains accuracy of above 0.711 even when detecting photos with various random noise. Compared to Biformer, PCBT reduced the parameter count by 22.4%. Additionally, it achieved improvements in recognition accuracy on the noiseless validation sets for cucumber, banana, and grape by 3.2%, 10.8%, and 1.5%, respectively. On the validation sets with 0-200 random pixel noise, the recognition accuracy also increased by 13.2%, 7.8%, and 17.3%, respectively. When compared to other lightweight models mentioned in the experiments, such as MobileNet, EfficientNet, and MobileFormer, PCBT demonstrates superior robustness. Furthermore, in comparison to more robust models like Swin Transformer, ConvNeXtV2, and DeepViT, PCBT not only maintains excellent robustness but also has fewer model parameters and lower FLOPs. Our proposed model aligns better with the practical requirements of agricultural applications.

Plant phenotyping relevance

作物画像から病害を識別する新規Transformerモデルを開発し、複数データセットおよびノイズ条件で精度・頑健性を比較検証しており、植物病害状態の画像ベース表現型取得が中心である。

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.