Abstract: RGB-D indoor scene classification is a challenging task. In this field， convolutional neural network has yielded excellent outcomes in terms of scene classification. However， many problems arise in the immediate application of traditional convolutional neural networks to indoor scene classification due to the multiple objectives， complex layout of indoor scenes， and the similarity existed between different categories of scenes. Aiming at these problems， an improved RGB-D indoor scene classification method based on convolutional neural networks is proposed， including two branches， one of which is a global feature extraction branch based on ResNet-18 and the other is a fusion branch of depth and semantic information. The features obtained from the two branches are fused for the purpose of indoor scene classification. Experimental results based on the SUN RGB-D dataset have proven the superiority of the proposed method in contrast to existing comparison methods.

Key words: convolutional neural network, scene classification, deep learning