An Adaptive Differentiable Neural Network Architecture Search Algorithm and Its Application for sewer defect detection

Differentiable neural network architecture search is now a popular way to automatically build deep networks in the past few years, thanks to its ease of implementation and efficiency. However, gradient-based search methods within differentiable spaces often suffer from search bias, particularly a tendency to favor skip connections, which can lead to suboptimal network performance. Although existing studies attempt to mitigate this issue by improving the exploration capacity of the algorithm, these approaches often incur significant computational overhead. In this work, we propose an adaptive exploration-based differentiable neural network architecture search algorithm, named AE-DARTS. AE-DARTS improves the effectiveness of algorithm exploration from both search and computation perspectives. In the search-direction-oriented approach, a composite interference term is introduced to enable both effectiveness and randomness in exploration. On the computational side, a dynamic partial edge sampling search strategy is proposed to reduce exploration costs and improve the exploitation precision. Furthermore, we have reimagined the cell structure to satisfy the lightweight demands of the AE-DARTS application for detecting sewer defects. Experiments show that the proposed method can find neural network architectures with higher feature representation ability with the same or even less computational overhead than current differentiable neural architecture search methods.