Deep learning for underwater image recognition in small sample size situations

Underwater target recognition is a challenging task due to the unrestricted environment of the ocean. With large datasets, deep learning methods have been applied with great success to the image recognition of objects in the air. However, it has been observed that deep neural networks (DNNs) easily suffer from overfitting with small samples. Underwater image acquisition always requires much manpower and costs a lot, which makes it difficult to obtain enough sample images for training DNNs. Besides, images captured by underwater cameras are usually deteriorated by noise. Taking live fish recognition as an example, we proposed a framework for underwater image recognition in small sample size situations. First, a novel improved median filter was utilized to suppress noise of fish images. Then, a convolutional neural network was employed and pre-trained with images from the world's largest image recognition database-ImageNet. Finally, preprocessed fish images were used to fine tune the pre-trained neural network and test the classification performance. Experimental results showed that the approach is capable of recognizing fish species, which provides an effective way for solving recognition tasks in small sample size situations.