ADNet: An Attention Embedded DenseNet121 Model for Weed Classification

Weed management is a critical factor that significantly impacts crop yield and overall crop productivity.Identification of the appropriate weed has been the big problem for the farmers over the few decades since weeds have the similar appearances with crops. This paper presents the vision-based weed classification system using convolutional neural networks and attention mechanism for maize crop. A model named ADNet has been presented in this study by employing DenseNet121 model with convolutional block attention mechanism (CBAM). ADNet yields the good performance with 99.58% accuracy for weed identification in maize crop. The proposed work was implemented with image dataset of maize crop with 6000 images and four weed species of corn crop. CBAM has been used on the feature map of DenseNet121 for extracting the distinguish features and focussing on the local regions. In addition, some well-known CNN models like EfficientNetB0, DenseNet121, MobileNetV2, InceptionResNetV2 and ResNet50V2 were also implemented in this work and their performance were compared with proposed work. Amongst all experimented CNN models, our proposed model ADNet (DenseNet121 + CBAM) outperformed all five CNN models implemented. The presented model shows the effectiveness of the attention mechanism embedded DenseNet121 model to significantly distinguish and classify the weeds in the maize crop.