Application of Deep Learning to Underwater Invasion Warning System

In this paper, we discuss our trying to improve detection performance of a diver detection sonar using deep learning in very shallow water area. We had developed and evaluated our underwater invasion warning system to detect intruders, diver and diver with swimmer delivery Vehicle (SDV), which try to intrude a harbor for several years. We find out that there are some demand for the underwater monitoring, especially in very shallow water area, less than ten meters depth. We have already known that such very shallow water makes difficult situation for sonar operation because reverberation is high and false alarms happen frequency than usual. We have already developed better way by analyzing detection parameters such as a signal level and a signal length by a statistical analytic method and solving the problem which uses an automatic optimization method of detection parameters. But in general, it is known that there is a trade-off relation between the reducing false alarms and good detection performance. There may be case that it is difficult to combine good reducing false alarms and good detection performance by a statistical analytic method. Therefore, we have worked on development of the method to solve the problem. In this work, we discuss the detection method using deep learning as one of our trying. Like general surveillance camera, a background image will be extracted. And we will detect the moving object by subtract difference between the background image and the current image. In this paper, we developed a new algorithm to detect moving object by the same way, which employs the differences between i) the observed current plan position indication image in short PPI image and ii) the predicted PPI image generated by deep learning as a background image. The training for the predicted PPI image was executed including the reverberation in the sea. Our proposed algorithm becomes one of the new detection method which does not use the conventional method using comparison with threshold value to decrease the false detection rate.