Research on the Hyperparameter Optimization Method of Mask RCNN Based on DARTs

Abstract

When dealing with specific target detection tasks, the target detection algorithms based on deep learning are often due to the lack of training samples or expert experience, resulting in the training and verification of the network requires a lot of computing resources, and cannot achieve better detection performance. This paper proposes a hyperparameter optimization method and training method of Mask RCNN Based on DARTs. Experiments show that,
this method can optimize the model with only a small number of training samples and without expert experience, and improve the accuracy of the model by 4.83% when IoU is 0.75.