Building recognition is to obtain building image information through computer vision equipment,and then use building recognition technology to classify and identify the buildings in the image.This technology has extensive application value in the fields of intelligent video navigation,smart city intelligent transportation,and unmanned device positioning.The main research contents of this article are as follows:By analyzing the influence of noise such as non-uniform illumination or local occlusion on building recognition,this paper proposes a building recognition method based on the neighborhood sensitive gradient histogram gist feature.This method is different from the traditional method of extracting texture features based on the Gabor filter,which uses the neighborhood sensitive gradient direction histogram of the building image as the texture feature of the image.The neighborhood sensitive gradient direction histogram of each pixel is a multi-directional histogram based on the entire building image.Therefore,our nearest neighbor gradient histogram gist feature has strong texture description ability,and is not sensitive to noise such as non-uniform lighting or local occlusion.Based on the neighborhood sensitive gradient direction histogram,this paper proposes a multi-scale neighborhood sensitive gradient direction histogram feature to extract the texture features of buildings.This method uses a three-layer image pyramid to extract texture features,which can well avoid the impact of building size on classification results.Finally,this paper proposes a building recognition method based on spatial texture and sparse representation of color features.This method uses multi-scale neighbor sensitive gradient direction histograms and color auto-correlation maps with spatial information to extract the spatial texture and spatial color features of building images.The extracted building color,texture and lightness features are combined to form joint features.The joint features are sparse representation and dimension reduction by self encoder,and are classified by limit learning machine.The methods proposed in this paper have been tested many times on the Sheffield buildings dataset.And the average recognition rate of the proposed method is over 10%higher than that of GIST feature,which proves the effectiveness of the proposed method. |