Scene Identification Based On Computer Vision

Scene identification belongs to the category of target recognition and has broad application prospects in the field of computer vision applications.Scene identification is similar to face recognition which belongs to the "one-to-many" pattern recognition problem,but the challenge is that the scene image imaging environment is complex.Due to uncontrollable factors,there are sharp changes in the scale,perspective and lighting of the image.It may also because the time span is too large and the background changes significantly,moreover there are disturbance factors such as seasons,weather,and moving targets.Therefore,scene identification is an under non-controlled natural environment with complex and dynamic changes in computer vision.These problems have not yet been well resolved.This article studies several scene identification methods from different perspectives.The main work is as follows:(1)Based on the artificial feature extraction technology,the scene identification is achieved through the principle of local feature point matching.Aiming at the problem of serious mismatch in traditional single-point image matching technology,a fast scene identification method based on LDB descriptor and local spatial structure matching is proposed.This method inherited the fast and economical storage space characteristics of the binary feature descriptors.It replaced single-point matching with multi-point matching to eliminate a large number of mismatches,and improved the matching accuracy.This method achieved a good balance between real-time performance and robustness.It has achieved better results than the traditional matching method in the landmark building data of Xiangtan University.(2)Regarding scene identification as an object classification problem,and using Deep Neural Network to Realize Scene identification.In-depth network through hierarchical feature learning,free from the limitations of the traditional artificial extraction features.This article used the AlexNet network to learn and test the extended Xiangtan University landmark building data set,verifying that this method greatly improved the success of scene recognition.The accuracy and reliability were better than the feature point matching methods.(3)Scene identification based on discriminative feature expression learning.In order to make the image feature expression of learning more discriminative,this paper studied the scene identification method based on the Siamese convolutional neural network.Through the characteristic expression learning of the degenerate convolutional neural network,the image pairs of the same scene are distributed in the feature space compactly,and the images of different scenes are spaced far apart in the feature space.The method was tested on publicly available Street View datasets with dramatic changes in lighting characteristics.Under the premise that only a few pairs of training samples are required,the Siamese convolutional neural network achieved excellent learning effects and well resolved lighting changes.Under the scene identification problem,the recognition effect was superior to the method of using in-depth network classification to solve this kind of scene identification problem.The scene identification methods studied in this paper have tried many aspects in solving the problem of dramatic changes in the scenes of scale,perspective,and lighting.The results obtained have positive implications for reference.The relevant methods are also worth further research and analysis.