Improving Deep Learning-based Plant Disease Classification with Attention Mechanism

Abstract In recent years, deep learning-based plant disease classification has been widely developed. However, it is challenging to collect sufficient annotated image data effectively train learning models for recognition. The attention mechanism in assists the model focus on informative segments and extract discriminative features of inputs enhance training performance. This paper investigates Convolutional Block Attention Module (CBAM) improve with CNNs, which a lightweight module that can be plugged into any CNN architecture negligible overhead. Specifically, CBAM applied output feature map CNNs highlight important local regions more features. Well-known (i.e. EfficientNetB0, MobileNetV2, ResNet50, InceptionV3, VGG19) were do transfer then fine-tuned by publicly available dataset foliar diseases pear trees called DiaMOS Plant. Amongst others, this contains 3006 images leaves affected different stress symptoms. Among tested EfficientNetB0 shown best EfficientNetB0+CBAM outperformed obtained 86.89% accuracy. Experimental results show effectiveness recognition accuracy pre-trained when there are few data.