Scene Recognition with Convolutional Residual Features via Deep Forest
Mingfei Han, Shengyang Li, Xue Wan, Guiyang Liu · 2018
Convolutional Neural Networks (CNNs) have made remarkable progress on image classification and other relative computer vision filed, which need large-scale data for training. In this paper, a method named DFCRF (Deep Forest with Convolutional Residual Features) is proposed. It is based on the gcForest proposed by Zhou and Feng. And we use a recent released AI Challenger dataset, containing around only 50,000 images mainly captured in China. Different from utilizing only CNNs, we use convolutional residual features for further recognition, followed by gradient-based XGBoost and cascade deep forest. Then, we conduct extensive experiments on the AI Challenger dataset and reconstructed Places2 dataset to show the effectiveness of our method.