Research On The Fast Scene Classification Based On Gist Of A Scene

Visual is the most important method to perceive environment. It is proved that80percents of the perceived information is from human eyes. As a result, the research ofcomputer visual has become a hotspot in artificial intelligence field. And due to fastdevelopment of computer science, the quantity of visual/video databases is increasing at ahigh velocity as well. Automatic labeling and classification methods are in research to replacetraditional manual methods.Scene classification is an important branch of visual classification. The conceptions ofscene description and scene understanding are specified in the2006MIT Workshop on SceneUnderstanding, and they are regarded having a bright prospect. Three respects are included inthe scene classification thinking: scene classification based on object, scene classificationbased on region, and scene classification based on context.The main research of this paper is fast scene classification based on scene Gist. It is afurther study of scene classification based on context. Scene Gist classification is widely usedin recent years. Object identification and region partition steps are omitted, compared withtraditional classification methods. The scene visual is regarded as an entirety to implementobject extraction and classification.Based on this background, three respects of work are included in this paper:(1) The spatial envelope model is studied. It is an early scene classification model.Unlike most models that directly implement eigenvalue extraction algorithm in original visual,the spatial envelope model extract degree of the naturalness, openness, roughness andexpansion under Fourier frequency spectrum. In addition, discriminant spectral template forthe degree of naturalness is implemented based on spatial envelope model. So the naturalnessof a real scene is calculated from this template, and used as the fundamental characteristic ofnature-man-made scene classification.(2) The Gist model based on the visual cortex feature channel is improved in the paper.The Gist model based on visual cortex feature channel is propose according to the Itti saliencymodel, which is a branch model of the scene recognition model proposed by Itti and Siagian.And Its basic idea is that the gist vector of a scene image is extracted from the features maps(color feature maps, intensity feature maps and orientation feature maps) of the visual cortexfeature channel. The Gist model is improved because the shortcoming is analyzed accordingto its experimental results, so the Gist model based on visual cortex feature channel and thedegree of naturalness is proposed. The degree of naturalness is added to the Gist model basedon the visual cortex feature channel. And the scene database is divided into the natural scene and the man-made scene before the gist vector is extracted. Then the semanticclassification is further done in the two categories.(3) A new model is implemented in this paper：the Gist model based on visual cortexfeature channel and the degree of naturalness；And some comparative experiments which areused to verify that the new model provides a stronger classification ability are designed.Firstly, four sets of comparative experiments with the degree of naturalness are made in theSection5，which verifies that scene is well classified as natural and man-made scene by thedegree of naturalness. And what conditions can impact on classification results is analyzed inthe paper. Then two sets of comparative experiments are made for the new model, and arecompared with some existing scene classification model and algorithms. And it is proved thatthe new model provides a stronger classification ability.The spatial envelope model and the Gist model based on visual cortex feature channelare intensively studied in the paper. The Gist model based on visual Cortex Feature channelis improved because of its shortcomings. And the new model called the Gist model based onvisual cortex feature channel and the degree of naturalness is proposed in the paper. And it isproved that the new model provides a stronger classification ability.