Pest Recognition Method Fusion Dilated Dual Attention and CapsNet Based on U-Net
Jiashan Liu; Zhiwen Ren; Xiaoli Shi; Zili Ding
Convolutional neural networks (CNN) have shown remarkable success in various computer vision tasks, including pest recognition. However, U-Net, a popular CNN architecture for image segmentation, lacks position and spatial information. To address this issue, we propose a novel pest recognition method, called DACapsU, which fuses dilated dual attention and CapsNet with U-Net architecture. By replacing the fifth layer CNN with CapsNet, DACapsU obtains more abundant position and spatial feature information. Additionally, the dilated dual attention module is integrated into the skip connection structure to acquire context information, space, and channel informati- -on. Furthermore, a global information module is designed in the image input stage to obtain global information among datasets. We evaluate the proposed method on a rice pest image dataset with complex backgrounds, achieving a recognition accuracy of 94.45%. Compared with VGG16 and ResNet, DACapsU improves the recognition accuracy by 5.51% and 2.89%, respectively. DACapsU can identify pests with different sizes, complex backgrounds, and diverse postures, making it a promising candidate for pest recognition systems. Our proposed method provides a new approach to address the issue of lacking position and spatial information in CNN.