Image aesthetic evaluation using paralleled deep convolution neural network

TL;DR

This paper introduces a parallel deep convolutional neural network architecture for automatic image aesthetic evaluation, addressing overfitting and feature extraction efficiency issues in traditional DCNNs.

Contribution

The paper proposes a novel parallel multi-level CNN architecture that adapts to training datasets, improving aesthetic feature extraction and overall evaluation performance.

Findings

PDCNN outperforms traditional methods in aesthetic evaluation.

The architecture effectively balances complexity and feature extraction.

Experimental results demonstrate improved accuracy.

Abstract

Image aesthetic evaluation has attracted much attention in recent years. Image aesthetic evaluation methods heavily depend on the effective aesthetic feature. Traditional meth-ods always extract hand-crafted features. However, these hand-crafted features are always designed to adapt particu-lar datasets, and extraction of them needs special design. Rather than extracting hand-crafted features, an automati-cally learn of aesthetic features based on deep convolutional neural network (DCNN) is first adopt in this paper. As we all know, when the training dataset is given, the DCNN architecture with high complexity may meet the over-fitting problem. On the other side, the DCNN architecture with low complexity would not efficiently extract effective features. For these reasons, we further propose a paralleled convolutional neural network (PDCNN) with multi-level structures to automatically adapt to the training dataset. Experimental results show that our proposed PDCNN architecture achieves better performance than other traditional methods.