Research On Image Scene Classification Algorithm Based On Global And Local Information

Scene classification is an open and challenging issue in image understanding and computer vision.Effective image representation is gaining increasing attention in scene classification tasks.This paper aims to construct feature extraction algorithms to form a more abundant and effective image description,so as to achieve high-precision classification of scene images.First,an image scene classification algorithm based on multi-level feature representation is constructed.The traditional scene classification uses a single low-level feature to construct an image description,which cannot effectively represent the scene image with variable content.Therefore,the dense scale invariant feature transform(SIFT)features and the convolutional layer's convolutional neural networks(CNN)features of the sampled image block are extracted,the local features of the image block are respectively encoded by the vector of locally aggregated descriptors(VLAD)method,and the low-level image description and the middle layer image description including the local semantic information are constructed.At the same time,the low-level description and the middle-level description of the image are merged into the high-level semantics of the fully connected layer of the image,thereby obtaining an accurate image representation that integrates the local spatial information and the global semantic information.Secondly,an image scene classification algorithm based on local feature coding and multi-channel feature fusion is designed.In the common algorithms,the full connection layer feature is used to represent the image,and the local information of the image is lacked,which reduces the classification and discriminating ability.Therefore,by analyzing the feature performance of different channels of convolutional neural networks,the CNN features of different channels are fully utilized to complement each other.The locality constrained linear coding(LLC)method is used to encode the convolutional layer CNN features of multi-size local image blocks to obtain local information of the image.The coding features are multi-channel fused with the global CNN features of the fully connected layer of the original image to obtain a more efficient image representation.Finally,an image scene classification algorithm based on discriminative clustering and weighted description is proposed.In the traditional scene classification,the K-means method is used to cluster the underlying features to construct the visual codebook.The clustering effect is sensitive to the size of the codebook and the initial clustering center,and the underlying features lack semantic information and cannot effectively represent the image.Therefore,a discriminative clustering method is proposed.Using the correlation distance,the correlation distance matrix of each category of image features are subjected to secondary traversal clustering,and the feature mean of each cluster cluster is taken as the cluster center to construct a visual universal codebook.After the image is segmented,the local description of the image is obtained based on the general visual codebook mapping,and the global description based on the image depth CNN features are combined and weighted to obtain a richer image representation.