Research On Scene Classification Based On Features’ Dimension Reduction

Scene classification annotates images automatically based on a groupof given semantic labels, which is very important in object recognition,content-based image retrieval, image filtering and image enhancement.However, due to variation of illumination, scale, rotation, angle and pose,it is still very difficult to classify the scene images correctly. Nowadays,scene classification has become an important and challenging issue in thefield of digital image processing.This paper focuses on scene classification and proposes a method thatreducing dimension basing on visual words of SIFT represented by spatialpyramid. Then I study on fusing features after reducing their dimensionseparately.The study reveals that, image features always contain some redundantinformation or noise. This will not only increase complexity of learningclassifier, but also affect the accuracy and efficiency of the classification.Therefore, we propose to reduce the dimension of the features in the sceneclassification.This paper first proposed a method based on dimension reduction ofSIFT’s visual words. We divide image into several patches, and extractfeature on every patch. In this paper we extract feature on three differentscales. Then we divide image into several blocks according to spatialpyramid. Then we use bag-of-words, sparse coding, max pooling andspatial pyramid to obtain feature of each block. We reduce dimension onevery visual word of all blocks. At last we get the representation of wholeimage. Experimental results show that, reducing the dimension of visualwords after representing image using spatial pyramid, not only adds spatial information of image, but also eliminates redundant information and noise,and reduce the dimension of the input vector of the classifier to makeclassification learning better, therefore increasing the average recognitionaccuracy rate of the scene datasets. On Scene-8datasets, our method’saverage recognition accuracy achieves89.5%. On Scene-15datasets, ourmethod’s average recognition accuracy achieves84.0%,3.0-3.7%higherthan other methods.Subsequently I studied a method based on reducing dimension ofseveral features and then fusing them. When we reduce dimensions offeatures’ visual words firstly, then assign different weights to differentfeature, we can represent the image information better, and can improvethe accuracy of scene classification. The experimental results show that, onScene-8datasets, our method’s average recognition accuracy achieves90.5%. On Scene-15datasets, our method’s average recognition accuracyachieves86.5%,3.0%higher than other methods.