ILGNet: inception modules with connected local and global features for efficient image aesthetic quality classification using domain adaptation

Abstract

In this paper, we address a challenging problem of aesthetic image classification, which is to label an input image as high or low aesthetic quality. We take both the local and global features of images into consideration. A novel deep convolutional neural network named ILGNet is proposed, which combines both the Inception modules and an connected layer of both Local and Global features. The ILGnet is based on GoogLeNet. Thus, it is easy to use a pre-trained GoogLeNet for large-scale image classification problem and fine tune our connected layers on an large scale database of aesthetic related images: AVA, i.e. domain adaptation. The experiments reveal that our model achieves the state of the arts in AVA database. Both the training and testing speeds of our model are higher than those of the original GoogLeNet.

Transfer learning of ImageNet Object Classification Challenge features to image aesthetic binary classification

“…Another important model used in both binary and multiclass aesthetic classification are neural networks. A fully connected network can be used with softmax activation in the last layer for classification algorithms [Ma et al 2017, Lu et al 2015, Talebi and Milanfar 2018, Jin et al 2016b].…”

“…It is also possible to obtain generic features by using transfer learning on DCNNs that have been previously trained to classify objects, like Inception, VGG and MobileNet [Talebi and Milanfar 2018]. These transfer learned descriptors can also be used with additional convolutional layers [Jin et al 2016b].…”

Two open-source projects for image aesthetic quality assessment

“…Compared with previous methods, the classification accuracy is greatly improved. The accuracy we achieve on the AVA dataset is 81.68%, and the accuracy is up to 82.66% by using the Inception V4 module [1,2].…”

Atrial Fibrillation Identification With PPG Signals Using a Combination of Time-Frequency Analysis and Deep Learning

“…Many well-known deep learning models, such as VGGNet [37], GoogLeNet [38], and Microsoft ResNet [39], have achieved excellent results in international classification recognition competitions, and they offer many ways to solve problems. In this study, we compare the proposed 2-CLSTM Model with VGGNet, GoogLeNet, and ResNet50 using the same input data set and Adam [40] optimizer.…”

“…The Flatten structure is used to connect the largest pool layer to the fully connected layer.Compared to VGGNet, GoogLeNet has a lighter structural design. Inception [38] is the core structure of GoogLeNet. This structure stacks convolution and pooling operations in CNN, which allows the model to increase both network width and network adaptability to scale.…”