Hierarchical Encoding Method for Retinal Segmentation Evolutionary Architecture Search

Evolutionary computation (EC) based method for Neural Architecture Search (NAS) is a thriving research field. Current NAS research typically focuses on concurrently searching for architecture and given operational hyperparameters (e.g. kernel size, stride, and padding). The architecture and its operational hyperparameters obtained through this method may not necessarily be the most optimal match, ignoring the exploration of operational hyperparameters space. To address this challenge, we propose an EC-based NAS method, namely EA-FCNet, to strike a balance between global search and local search. Specifically, we design a hierarchical encoding strategy distinguishing architecture and operational hyperparameters so that the algorithm can search both in a nested manner. Meanwhile, unlike existing methods based on convolutional neural networks (CNNs), our search space combines convolutional and fully connected operations to extract comprehensive features and enhance feature association. Furthermore, we address the challenge of combining these two operations that have mismatched input and output shapes by introducing a repairment strategy that allows the algorithm to handle such situations seamlessly. To evaluate the performance of the proposed algorithm, we conducted extensive comparison and ablation experiments on two publicly available datasets: DRIVE and CHASE_DB1, which demonstrate the effectiveness of the proposed algorithm, with the obtained architecture achieving competitive results. Additionally, the proposed updating strategy has increased the search speed tenfold.