Scene Recognition Based On Sparse Coding And Extreme Learning Machine

Scene recognition identify the images where they belongs by image processing technology and pattern recognition technology.It is one of the important studies in computer vision and robotics field,and promotes the development of image retrieval technology and intelligent robot navigation applications.Scene recognition is mainly composed of feature extraction and classifier recognition.This paper studies image feature extraction and recognition algorithm respectively,proposes two different feature extraction algorithms,and uses extreme learning machine(ELM)algorithm as recognition algorithm.The research contents an contributions are as follows:Two different extraction algorithms are proposed in the feature extraction process,one is the middle layer feature extraction algorithm based on the sparse coding technique,and the other is the feature extraction algorithm of the scene image based on the convolution neural network.The feature extraction algorithm based on sparse coding consists of three steps,which are Dense SIFT for low-level feature extraction,sparse coding and spatial pooling.Dense SIFT has the invariance on rotation,brightness change,but also has a certain degree of stability on the angle of view,affine transformation,noise,which improves the robustness of scene feature representation.Sparse coding can achieve a sparse representation of the data,thus improving the ability to distinguish features of expression.This paper not only adopts the feature-sign search algorithm based on L1 norm,but also proposes homotopy iterative hard thresholding method(HIHT)coding based on LO norm.The spatial pooling process divides the area of the image,and separately counts the middle-level features in the area,which can prevent the loss of spatial distribution information.This paper presents a spatial layout sensitive pooling method.The space layout for pooling is based on three rectangles with size of 1*1,1*4 and 4*1 in each image.They are derived from inherent characteristics of the scene images by regularly dividing the image in horizontal and vertical direction.The depth learning method based on convolution neural network can automatically studies sparse and multi-level feature.This paper uses ELM based on kernel function in scene recognition to solve the problem of highly nonlinearity of classification boundary,and obtain better recognition performance.The above-mentioned two scene recognition methods are verified in 15-scenes database respectively.The experimental results show that the recognition accuracy of the two feature extraction methods combined with ELM are 86.23%and 88.38%respectively.The first feature extraction algorithm based on sparse coding technology,whose process is simple,can directly extracte middle feature expression.The second feature extraction algorithm based on convolutional neural network is a complex network model,and needs to be trained repeatedly to find a better set of network parameters.Experimental results show that the ELM classifier has higher recognition accuracy than other classifiers.In summary,the proposed scene recognition method in this paper has better recognition performance than other mainstream scene recognition method.