Transfer Learning for Small-Scale Fish Image Classification

Abstract

Fish image classification on small-scale datasets is a classical fine-grained classification problem with more challenging than common classification problems since popular Convolutional Neural Networks (CNNs) always need massive labeled images to achieve best effects. This paper presents a method for fine-grained fish image classification on small-scale dataset by improving transfer learning with Bilinear Convolutional Neural Networks (B-CNNs). Contrast to the popular CNNs for image classification such as VGG net or ResNet, our method is capable of classifying images with insufficient training data. We evaluate the performance of our method and state-of-the-art CNNs on a small-scale fine-grained dataset (Croatian fish dataset) with only about 10 samples per category. The experimental results show that our transfer learning method can greatly enhance the accuracy on four popular CNNs. In particular, the method on Inception-v4 with 78.25