For unmanned systems used in offshore surveillance and coastal security, image recognition functions which can detect and identify targets and obstacles from camera images are essential. To achieve high image recognition, learning using images of scenes where accuracy of detection and identification is low is required. However, acquiring images for learning and searching for images with specific scenes requires significant effort. Therefore, a method to automatically search for images of specific scenes from large-scale image databases available online, using images with low detection and identification accuracy as input, has been developed. As a result, the workload required for this task could be reduced. Wide application of this developed technology to products equipped with image recognition functions that use deep learning is planned.
