Research On Semantic Annotation Of Traffic Scene Image

Semantic annotation of images is to group the image pixels according to the semantic meaning,which means that each pixel is assigned a category label.This can complete the image segmentation and multi-target recognition tasks simultaneously.For example,individual machine vision problems such as pedestrian detection,road segmentation and object classification can be integrated in a unified framework by semantic annotation methods,which can achieve overall scene understanding.Therefore,semantic annotation has important research significance in the field of machine vision.This paper focuses on studying nonparametric semantic annotation algorithm for traffic scene images.Improved the three main steps of nonparametric semantic annotation: image retrieval,superpixel matching and MRF labeling.The main research works are listed as follows:1.Aiming at the problem of using random sampling to calculate weights in weighted hash code ranking algorithm,which leads to inaccurate weight assignment and low retrieval accuracy,a coarse-to-fine hash code weighted ranking image retrieval algorithm is proposed.Firstly,a shorter hash code is generated to improve coding efficiency.Then the weight value of each hash bit is calculated in the sampling subset,which can be generated based on data dependent dissimilarity,and a candidate nearest neighbor set is obtained according to the weighted Hamming distance.Finally,the ranking score of the data is calculated in the set,the retrieval accuracy is further improved by reordering according to the score,and the nearest neighbor search of the query image is realized.The experimental results show that the mean average precision of the proposed algorithm increased by 13.33% and 11.61% respectively when the coding length are 48 bits and 96 bits on MNIST database.2.Nonparametric semantic annotation is easily affected by the accuracy of image retrieval and unbalanced dataset,which caused inaccurate semantic annotation.To solve these problems,a scene semantic annotation algorithm based on CNN feature and improved superpixel matching is proposed.Image features are obtained through convolutional neural network then reduce dimensions of features,which can improve the accuracy of image retrieval.Superpixels of the images in the retrieval set are weighted by using Gaussian kernel density estimation,which can improve the superpixel matching accuracy of rare classes.Therefore,semantic annotation accuracy of query image can be improved.The experimental results on SIFT flow and KITTI dataset show that compared with state-of-the-art methods,both per-pixel and per-class rates are the best.3.The basic nonparametric framework is applied to road scene parsing without taking scene structure into account,which make objects label error and rare categories labeling inaccurate.To improving the basic method,a structure prior for nonparametric road scene semantic annotation algorithm is proposed.Using multi-scale superpixel segmentation to improve matching accuracy.Structure prior histogram of every category is computed by stable rode scene layout.Adding this information into data term of MRF energy function can improving labeling accuracy.The experimental results show that compared with original nonparametric method,per-pixel rate and per-class rate of the proposed algorithm raises 3.3% and 5.1% respectively on KITTI dataset.