Biometric Fish Classification of Nordic Species Using Convolutional Neural Network with Squeeze-and-Excitation

Abstract

Squeeze-and-Excitation (SE) is a technique within convolutional neural networks (CNN) that can be applied to existing CNNs by applying fullyconnected layers between convolutional layers and merging the outputs. SE was the winning architecture of the ImageNet Large Scale Visual Recognition Challenge (ILSVRC) in 2017. In this thesis, we propose a CNN using the SE architecture for classifying images of sh. Previous work in the eld relies on applying lters to the images to separate the sh from the background or sharpen the images by removing background noise. The images from the dataset are extracted from underwater cameras and contain noise, which is why classifying these images is challenging. Di erent from conventional schemes, this approach is divided into two classi cation problems. The rst approach is to classify sh from the Fish4Knowledge dataset without using image augmentation, and the second is to classify sh from a new dataset consisting of Nordic species. We name the rst approach pre-training, and the second post-training. The weights from pre-training are applied to post-training. Our solution achieves the state-of-the-art accuracy of 99.27% accuracy on the pre-training. The accuracy on the post-training is lower with an accuracy of 83.68%. Experiments on the post-training with image augmentation yields an accuracy of 87.74%, indicating that the solution is viable with a larger dataset. Keywords: Classi cation, CNN, Squeeze-and-Excitation