Research On Indoor Scene Recognition Via RGB-D Images Based On Convolutional Neural Networks

Scene recognition is an important subject of computer vision,and is widely applied in many areas.These areas mainly include information retrieval in large scale image database,the mobile positioning and interacting with environment of robotics,the event monitoring in the field of security monitoring,etc.Scene recognition has significant research and application values.Indoor scene recognition is more challenging compared to outdoor scene recognition,which is determined by the complexity of indoor scenes.The depth information can effectively improve the performance of scene recognition.Convolutional neural networks has been proven to have excellent performance in large scale image recognition.The traditional computer vision methods for indoor scene recognition have low accuracy and poor robustness,and the manually extracted features is not generalized.In view of the shortcomings of traditional methods,this paper proposed to use CNNs models to research on indoor scene recognition via RGB-D indoor scene images.Firstly,we have researched the background and significance of indoor scene recognition,and analyzed technical difficulties and challenges.Then we study the feature selection and extraction of scene images,and analyze the advantages and disadvantages of traditional computer vision methods and convolutional neural networks.Based on AlexNet,we design 5 different CNNs models for different types of inputs,use Adam method to train our models.Though experiments,we analyze the performance of our CNNs models and the effect of adding depth information.To explore the mechanism of CNNs models,we visualize the weights.Besides,we use a technique to compute a class saliency map,specific to a given scene image and class.We observe the salience maps to study the attentions of scene recognition.The experiments show that the CNNs models have higher accuracies than the traditional computer vision methods,and adding depth information can improve accuracies.The visualization of CNNs models and the class salience maps can further help us design better CNNs architectures and select appropriate features for scene recognition.