Enhancing CNN Architecture with Constrained NAS for Boat Detection in Aerial Images

This article describes a Network Architecture Search (NAS) method based on an evolutionary algorithm. It applies constraints in the search space to find architectures adapted for real-time edge detection. Using this method, we start from an initial parent, Tiny Yolov2, which has a mean Average Precision (mAP) of 68.24% in our dataset. The method obtains an architecture which achieves a high mAP of 71.89% in the same dataset. This architecture registers an augmentation of operations’ count by only 8.21% compared to Tiny Yolov2 while remaining within the time budget constraint of real-time processing of our Unmanned Aerial Vehicle (UAV) system. This constraint defines a minimal limit of the number of Frames Processed per Second (FPS). It is set to 20 FPS, and it was obtained based on current flight conditions and equipment configuration. Compared to Yolov2, this same architecture achieves a gain in operations count of 80.28% while being near of the Yolov2 mAP. Combining two variants of our search method, we find another architecture that outperforms Yolov2 in accuracy and speed even if this architecture is trained with fewer iterations than Yolov2. This work demonstrates the high value provided by the evolutionary algorithm that we designed, the direct effect of the fitness function on the outcome, and the unexpected importance of retraining the parent on each new generation.