Indoor Scene Recognition Based On Deep Learning And Sparse Representation

With the development of computer vision and intelligent robot,scene recognition,as the key technologies,has become an important research topic in the field of machine learning.At the same time,it is also a difficult point in the field of scene recognition.If we can effectively improve the performance of indoor scene recognition,we will greatly promote the development of intelligent robots,image retrieval,video retrieval and other fields,and bring huge economic benefits.Therefore,the study of indoor scene recognition is very meaningful and challenging.Compared to the outdoor scene images,the content elements in the indoor scene images are more complex.Because of the diversity and complexity of the indoor scene images,in the same scene,the differences between the captured scene images are also larger.In addition,the scene images are also vulnerable to the interference of shooting conditions.Although collecting the same scene image,the change of scene image may also happen due to the change of acquisition device,occlusion,illumination and angle.Therefore,the image samples in the indoor scene recognition often show the characteristics of small inter-class variety and large intra-class variety.This makes it difficult to effectively represent the semantic information of the indoor scene by the traditional shallow feature learning,so the recognition performance is not good.For this reason,relevant researchers have made unremitting efforts to solve these problems,but still have not been able to solve the problem of low indoor scene recognition rate.The reason is that it is difficult to describe the semantic information of indoor scenes by using traditional shallow features,and can not effectively solve the semantic gap between the underlying features and high-level semantics.For this purpose,this paper first attempts to combine depth learning and sparse representation for indoor scene recognition.In terms of feature extraction,this paper improved the traditional bag-of-words model and represented a mid-level feature building algorithm.Fast R-CNN based multi-class detector is training for extracting object information to be as the low-level features.An improved bag-of-words model is designed to build mid-level features from object-based low-level features,which retain the spatial information of object-based low-level features.For improving the robustness of the proposed method,sparse representation is used to make the final decision of indoor scene recognition from mid-level features,where the sparse coefficients computed by?"-norm instead of?_#-norm.The work of this paper is as follows:(1)This paper briefly introduces the research background and significance of indoor scene recognition and analyzes the research status of indoor scene recognition by referring to a large number of domestic and foreign literature.It also introduces some public scene recognition datasets.(2)The typical feature extraction algorithms and classification algorithms used in typical indoor scene recognition are investigated.The recognition rate of the classical scene recognition algorithm system is compared,and the advantages and disadvantages of each system are analyzed.(3)A middle-level feature construction algorithm based on deep learning is proposed.Because of indoor scene images have the characteristics of small inter-class variety and large intra-class variety,this makes it difficult to effectively represent the semantic information of the indoor scene by the traditional shallow feature learning.It also means that it can not construct strong robustness,anti-interference and semantic middle-level features,so the recognition rate is low.In this paper,a middle-level feature construction algorithm based on deep learning is proposed.Firstly,we use the Fast-RCNN algorithm based on deep learning to detect the indoor scene images quickly and accurately,and the target detection information are used as the underlying feature.Then this paper improves the traditional bag-of-words model,low-level features extracted by Fast-RCNN algorithm to construct with spatial information,strong anti-interference ability and strong robust middle-level features.The bag-of-words model are introduced briefly in chapter 2 and the performance of the algorithm are proved in chapter 4.(4)A scene recognition method based on sparse representation is proposed.In view of the successful application of the SRC algorithm based on sparse representation in the field of face recognition,and the algorithm has robustness and strong anti-interference ability when the images have the problems of angle pollution,occlusion pollution,illumination change,etc.so this paper introduces the sparse representation algorithm,and the validity of this algorithm has been proved through many experiments in chapter 4,the SRC algorithm based on sparse representation are introduced briefly in chapter 2.(5)Summing up my work.At the end of the paper,I have made a summary of my work and put forward some places to continue to improve,and look forward to the research part of his own.