The application of deep neural networks in industrial inspection is limited due to the lack of instance-level labels. To address this issue, the work aims to propose a neural network algorithm that combines convolution and attention mechanisms under weakly supervised learning for practical surface defect detection on cardboard. By integrating channel attention modules and gradient-based activation mapping modules, this network enhanced the precision of class activation maps and realized the precise localization of cardboard surface defects. Additionally, a combination of inverted residual structures and upsampling layers was utilized to refine shallow features and improve the network's feature extraction capabilities, thereby accelerating the convergence speed. Experiments were carried out on the publicly available cardboard defect dataset, achieving classification accuracy and localization accuracy of 99.0% and 92.2% respectively under the training with image-level labels and demonstrating the effectiveness of the proposed algorithm. The disadvantages of a small number of instance-level labels and excessive subjectivity are avoided, which lays a foundation for the removal of defective cardboard based on robots.