Research On Superpixel Segmentation Algorithm Based On Feature Complementation

Superpixel segmentation is an important basic technology in computer vision.It is usually used as a preprocessing step in the image processing field.It has important applications in semantic segmentation,target detection and tracking,depth estimation and other fields.Superpixel segmentation methods are mainly divided into traditional machine learning based method and deep learning based method.The seeds-based superpixel segmentation method in traditional machine learning method realizes segmentation by analyzing the similarity between pixels and seed on the original image.This type of method has low computing resource consumption,and can flexibly adjust the position of the seed to achieve the segmentation result of specific requirements,but it is difficult to capture the semantic features of the image.The supervised based method based on deep learning methods relies on manually labeled task-specific semantic segmentation datasets to obtain deep features with good semantic properties,and then segment high-quality superpixels.However,due to the singleness of training constraints,As a result,the mapped depth features cannot be aligned with the image color features,which affects the generation of highly homogeneous superpixels by such methods.This thesis analyzes the key problems of these two types of superpixel segmentation methods,and proposes corresponding strategies to solve the above problems,thereby improving the quality of superpixel segmentation.(1)This thesis proposes a superpixel segmentation method based on vine spread model,which improves the ability of superpixels to segment original image details and twigs and balances the effect of superpixel adherence and regularity,and serves as the homogeneity complementary feature of the proposed deep learning method one of the sources.The method first extracts image color and gradient features to define a soil model,which establishes a"soil" environment for vines;then defines a vine state model by simulating the"physiological" state of vines.Aiming at the randomness and coarse perceptuality of the seed initialization of the seed based method,this method proposes a new seed initialization strategy,which perceives the image gradient at the pixel level without randomness.For the pixel assignment problem of the seed point method,we define a three-stage "parallel spreading" vine spreading process as a new pixel assignment scheme:the proposed nonlinear velocity helps to form regular-shaped and well-compact superpixels;the crazy spreading mode and soil averaging strategies help to enhance superpixel boundary adherence.Since the proposed algorithm is effective in restoring image details and balancing superpixel adherence and regularity,this method will further enhance the performance and effect of many advanced visual tasks as a pre-processing process,and improve processing efficiency.(2)This thesis proposes a hybrid feature alignment neural network superpixel segmentation algorithm based on feature complementarity to solve the problem of feature misalignment and non-adaptive downsampling of feature in supervised deep learning methods,thereby improving the homogeneity of superpixels and segmentation quality.First,the method designs a feature alignment module,which enhances the semantic sensitivity and homogeneity of superpixels by introducing the supeipixel segmentation results based on the vine spreading method as complementary features,and aligning the deep network hybrid features by combining the semantically sensitive features provided by the semantic segmentation dataset.In addition,the method introduces a deformable seed module,which obtains a more reasonable adaptive sampling offset by predicting the position of the downsampling to generate a more representative initial seed,thus effectively enhancing the superpixel segmentation quality.