Adaptive Sparse Semantic Representation for Waste Electrical Appliance Recognition

Deep neural networks are important models in the task of recognition of waste electrical appliances. However, deep neural networks are nonlinear models, leading to the lack of self- interpretation of appliance recognition models as a critical issue. We propose an adaptive sparse semantic representation (ASSR) model to improve the self-interpretation capability of appliance recognition models. First, an adaptive sparse representation component is constructed to enhance the ability of the appliance recognition model for sparse feature extraction. Second, residual sparse blocks are designed to ensure reliable network architecture and enhance the semantic representation ability of the appliance recognition model. Finally, the appliance recognition model obtains an interpretable network architecture by computing the optimal sparse solution, ensuring the strong self-interpretability of the recognition model. The experimental results show that the recognition accuracy of ASSR is improved compared with the traditional deep neural network, and ASSR has linear self- interpretation capability.