Data augmentation using image translation for underwater sonar image segmentation

Abstract

In underwater environment, the study of object recognition is an important basis for implementing an underwater unmanned vessel. For this purpose, abundant experimental data to train deep learning model is required. However, it is very difficult to obtain these data because the underwater experiment itself is very limited in terms of preparation time and resources. In this study, the image transformation model, Pix2Pix is utilized to generate data similar to experimental one obtained by our ROV named SPARUS between the pool and reservoir. These generated data are applied to train the other deep learning model, FCN for a pixel segmentation of images. The original sonar image and its mask image have to be prepared for all training data to train the image segmentation model and it takes a lot of effort to do it what if all training data are supposed to be real sonar images. Fortunately, this burden can be released here, for the pairs of mask image and synthesized sonar image are already consisted in the image transformation step. The validity of the proposed procedures is verified from the performance of the image segmentation result. In this study, when only real sonar images are used for training, the mean accuracy is 0.7525 and the mean IoU is 0.7275. When the both synthetic and real data is used for training, the mean accuracy is 0.81 and the mean IoU is 0.7225. Comparing the results, the performance of mean accuracy increase to 6%, performance of the mean IoU is similar value.