Dense Network Weed Classification Method by Fusing Attention Mechanism

Weeds significantly impact crop yields, making effective weed control crucial for improving agricultural productivity. Rapid and accurate weed identification is essential to enhance the efficiency of weed management. This paper introduces a weed identification method based on an improved DenseNet network. By incorporating the custom MSECANet multi-scale channel attention mechanism, the process extends to multiple scales more effectively to capture favorable features. Additionally, Depthwise Separable Convolution replaces some standard convolutions in dense networks, reducing model parameters while improving parameter utilization and training speed. The ReLU activation is replaced with the PReLU activation to prevent gradient vanishing when the input is negative during network training. The improved network was trained over 30 batches using the Cotton Weed Identification Dataset, achieving an accuracy rate of 98.20%, 3.75 percentage points higher than the original model. Compared to existing methods, this approach offers higher recognition accuracy and better classification performance for weed images.