Object recognition in forward-looking sonar images with Convolutional Neural Networks

Forward-looking sonars can provide high resolution images that can be used for different tasks in an underwater environment. However, image interpretation is still an open problem due to multiple issues inherent in acoustic imaging. In this work, we use Convolutional Neural Networks (CNN) for object recognition in forward-looking sonar images. We show that a CNN outperforms the state of the art for such kind of images by achieving an accuracy of 99.2%. While state of the art template matching methods have accuracies between 92.4% and 97.6%. We also compare the number of learnable parameters of CNNs and template matching that are required to achieve high performance. Our results show that CNNs require less parameters to provide better recognition capabilities that generalize well to unseen data.