Attention-Aware Based Differentiable Architecture Search Network for Waste Recognition

Manually designed deep neural networks have successfully forwarded waste recognition tasks in the resource recycling field. However, due to the diversity of waste samples, the feature extraction ability of inherently designed models fails to fully satisfy the requirements of real world applications. In this article, an attention-aware based differentiable architecture search (ADAS) network for waste recognition is proposed, which self-organizes to generate an optimal network structure according to diverse waste data. First, a structured search space with attention-aware modules is designed to enhance the diverse feature representations of waste data. Secondly, an efficient and differentiable structure search is achieved by continuously relaxing the representation of the network architecture and search space. Finally, the optimal architecture search process is evaluated by a bi-level optimization algorithm. Experimental results show that the proposed method achieves more satisfactory classification performances than the manually designed ResNet, DenseNet networks in the TrashNet dataset and the self-built household waste dataset.