Automatic deep learning for automatic target detection of sonar images

Abstract

The typical deep learning detection workflow of sonar images uses a transfer-learning paradigm, i.e., to select an external optical image pre-trained classification network as the backbone for feature extraction, and then to fine-tune the entire R-CNN network on detection datasets. Though transfer learning can effectively avoid overfitting and accelerate network training, transferred models are usually too redundant and cannot achieve the optimal generalization. This paper uses Automatic Deep Learning (AutoDL) to improve the automatic target detection performance of sonar images. An improved neural architecture search algorithm is applied to automatically design the architectures of R-CNN detectors. The automatically designed architectures don’t require external optical images to pre-train and have lower requirements for R-CNN parameters and computations. Experiments on an actual measured sonar image dataset composed of small targets show that, the automatic detectors designed by the neural architecture search algorithm are better than the transfer learning detectors in terms of recall, detection precision, computations, and parameters.