Image Aesthetic Assessment using Deep Learning for Automated Classification of Images into Appealing or Not-Appealing

摘要

Automated Image Aesthetic Assessment has been challenging to implement due to varied perceptions of people. This paper aims to tackle the matter and achieve better accuracy by adopting a deep learning neural network approach to perform image aesthetic classification. This research work presents a deep convolutional neural network framework that programmatically extracts high and low ranking features of an image and differentiates the dataset for analyzing areas of concern. Our model performs Image recognition using TensorFlow and Keras. A high-level network is employed to train and classify images. Additionally, the proposed model employs color contrast, depth of field, and rule of thirds to further improve the aesthetic performance of the model. This also uses GrabCut algorithm for interactive foreground extraction using OpenCV (Open Source Computer Vision Library). Our dataset, comprising 6000 images, is compiled from a range of sources online(Pinterest, Google, Flickr, Kaggle, Flickr) to make it as diverse as possible. Our experiments demonstrate that compared to traditional handcrafted models our Deep Convolutional Neural Network model yields significantly better categorization correctness (accuracy) of 73.27%. Thus, the Deep Learning Model helps exclusively to boost the performance of Aesthetic Assessment.