Research On Superpixel Segmentation And Fast Implementation Methods

As an effective image preprocessing method,superpixel segmentation is widely used in tasks such as target detection,image segmentation,and object tracking.Efficient and accurate superpixel segmentation is a challenging problem.The difficulty lies in two aspects: first,due to factors such as noise in images,large differences in the internal pixels of targets,and inconspicuous pixel differences between targets,the superpixels generated by the existing superpixel segmentation methods do not adhere well to the boundary of targets;second,with the update of image acquisition devices,the resolution of available images is getting higher and higher,which leads to the high computational complexity of the superpixel segmentation,which is difficult to meet the needs of some real-time applications.Based on previous works,this thesis focuses on how to improve the accuracy and speed of superpixel segmentation,and several novel methods for superpixel segmentation and implementation has been proposed.Experiments demonstrate the effectiveness of the proposed methods.The main research contents of this thesis are as follows:A superpixel segmentation method based on local Gaussian mixture model is proposed to improve the segmentation accuracy of superpixels.The method first uses Gaussian mixture distribution to describe the generation process of pixels.Each Gaussian distribution corresponds to a unique superpixel.After obtaining the estimated values of the model parameters,each pixel is assigned into a superpixel whose Gaussian distribution has the largest posteriori probability.Each superpixel is only allowed to appear in a local area associated with its label,reducing the number of pixels used to update the model parameters in the parameter estimation process.Each Gaussian function has the same weight in the Gaussian mixture model,which allows each superpixel to have a similar size expectation,which helps to generate similar-sized superpixels.In the parameter estimation process,the contradiction between the segmentation accuracy and the regularity is balanced by adjusting the lower limit of the eigenvalues of the covariance matrices.Experiments show that this method can generate regularity-controllable superpixels and exceeds eight superpixel segmentation methods which are selected for comparison in terms of segmentation accuracy.In order to improve the accuracy of video(sequence)superpixel segmentation,a superpixel segmentation method based on local Gaussian mixture model and sequence time correlation is proposed.Similar to the superpixel segmentation method based on the local Gaussian mixture model,each superpixel on each frame corresponds to a unique Gaussian distribution,and the method of maximizing a posterior probability is used to determine the labels of pixels.Generally,the color of the same target in two frames has less changes.Therefore,in order to make the superpixels with the same number on two consecutive frames belong to the same target as much as possible,this method uses the same color parameters for the superpixels with the same number.Because there are often moving targets in video sequences,different spatial parameters are used for the superpixels with the same number.Experiments show that this method outperforms,in terms of segmentation accuracy,four video superpixel segmentation methods which are the same kind.Aiming at the problem of large amount of calculation of linear spectral clustering superpixel segmentation method,a GPU-based parallel linear spectral clustering algorithm is proposed,which effectively improves the computational efficiency.The linear spectral clustering method updates the pixel labels from the perspective of superpixels,resulting in the label update process of each pixel not being executed in parallel.In order to solve this problem,this thesis proposes a pixel-centric label updating method,which gives linear spectral clustering methods the ability to fine-grain parallelism,making it suitable for GPU parallel architectures.Compared with the original method that runs on a 3.3 GHz CPU,this method achieves minimum speedups of 15.2x and maximum speedups of 29.8x on GTX TITAN X,which effectively improves the computational speed of the linear spectral clustering method.In order to improve the computational performance of the superpixel segmentation method based on the local Gaussian mixture model proposed in this thesis,a fast and parallel implementation algorithm of the local Gaussian mixture model based on GPU is proposed.According to our analysis,certain pixels require the same Gaussian distribution parameters when updating the auxiliary variables using the expectation maximization method.These pixels are assigned to the same thread block for processing in this thesis,and the parameters are loaded into shared memory at one time to reduce the number of accesses to global memory.There is a summation process during the update of each model parameter.Each of them is assigned a thread block,where each thread first calculates the partial summation of the parameters,and then uses the parallel reduction method to complete the parameter update.This parameter update method can effectively improve the utilization of the GPU.The experimental results show that compared with the CPU-based multi-thread parallel implementation method,the GPU-based implementation achieves minimum speedups of 21.3x and maximum speedups of 27.5x on GTX 1080,which effectively improves the calculation speed of the algorithm.In addition,in order to promote the application and development of superpixel segmentation algorithms,the experimental code of the four methods proposed in this thesis is open source,and the open source link is given at the end of each chapter.