An Adaptive Pooling Model for Aesthetic Quality Assessment Based on Convolutional Neural Network

Abstract

It is a requirement to find well-photographed and attractive photos from a large number of images, but image aesthetic quality assessment (IAQA) is challenging because of the subjectivity of aesthetic criteria. In this paper, we propose an adaptive pooling model for IAQA based on convolutional neural network (CAP) which consists of three parts: feature extraction, feature adaptive pooling, and obtaining quality scores. Through experimental comparison, we choose ResNet50 network as the image feature extraction network. In order to reduce the fixed input size limit, we introduce Spatial Pyramid Pooling (SPP) in model. The final score is weighted summation of the score distribution. The experimental results show that the proposed model achieves the state-of-the-art performance.