Deep-Sea Biological Image Augmentation: A Generative Adversarial Networks-Based Application

With the recent progress of convolutional neural networks in image processing and pattern recognition, deep learning has become the mainstream method of object recognition in images. Typically, A model is trained prior to inference and the training process requires many training samples with sufficient randomness. However, the detection and recognition of deep-sea objects for deep-sea video image data has always been limited by the size of the training samples in existing databases, current deep-sea image samples are not sufficient for supporting training deep neural networks for recognition purposes. With the current increasing usage of generating adversarial networks (GANs), this report introduces a GAN-based application to generate deep-sea image samples, and the improvements of training deep neural network models using augmented training samples.