Deep learning for multi-label scene classification

Abstract

Scene classification is an important topic in computer vision. For similar weather conditions, there are some obstacles for extracting features from outdoor images. In this thesis, I present a novel approach to classify cloudy and sunny weather images. Inspired by recent study of a deep convolutional neural network and the spatial pyramid matching, I generate a model based on the ImageNet dataset. Starting with parameters learned from other classification tasks, I fine-tune the model using outdoor images. Experiments demonstrate that our classifier can achieve state-of-the-art accuracy. Multi-label learning is a variant of supervised learning where the task is to predict a set of examples, which can belong to multiple classes. This is a variant of popular multi-class classification problems in which each sample has one class label only. It can apply to a wide range of applications, which include text categorisation, semantic image labelling etc.. A lot of research work has been done on multi-label learning with different approaches. In this thesis, I train a neural network from scratch based on the generated artificial images. The model is learned by minimising an error function based on the Hamming distance, through the backpropagation optimisation. The model has high capability of generalisation.