Steel Surface Defect Recognition Based on Evolutionary Network Architecture Search

Defect recognition algorithm in steel strip production lines can effectively improve the quality of steel products and promote the overall development of the industry. However, the Deep Neural Network (DNN)-based defect recognition methods exhibit sensitivity to target scale and texture information in specific scenes, making it difficult for generalization across datasets with significant differences. Therewithal, networks optimized for existing production environments require expert redesign when deployed in new contexts, which significantly increasing deployment costs. In this paper, we utilize evolutionary deep learning algorithms to automatically search for architectures that are optimally suited to specific steel strip surface defect datasets. By combining multiple indicators such as inference time, network scale, loss function value, and accuracy, we propose a multi-objective fitness evaluation function for edge device inference. To mitigate the computational intensity of Network Architecture Search (NAS), this research incorporates strategies such as search space optimization, predictive evaluation, and multi-GPU scheduling to expedite the process. The proposed algorithm enhances the cost-efficiency of network design and lowers the barriers to re-optimization of model structures when transitioning between production environments, thus facilitating practical implementation within the industry.