Research On Segmentation Of Similar Foreground Image Sets Based On Improved Random Walk And Transforming CNNs

Image segmentation is an important research in digital image processing and computer vision.Its task is to extract the foreground objects contained in the image,which lays a foundation for related visual tasks such as image retrieval and recognition.The existing single foreground segmentation algorith;m is usually based on user interaction,while the foreground segmentation algorithm of image set is mostly based on deep neural network that requires a large number of labeled image training.Although the related researches have made promising progress,due to the complexity and diversity of image con tent,performing foreground segmentation for an image set remains a challenging issue.This paper aims to study the task of segmenting a set of images with similar foregrounds.The basic strategy is to interactively segment a small number of sample images,and then construct Then construct a transforming-based convolutional neural network model to learn various pose transformations between similar foregrounds.Specifically,the main contents of this study include the following two aspects:(1)Random Walk image segmen tation algorithm combining local and non-local feature associationsThe existing interactive image segmentation algorithms based on graph cutting and Random Walk usually focus on the segmentation of single target foreground.When the image contains repeated foreground objects,these algorithms need to perform multiple Interaction marks on each target,and the segmentation effect is usually poor.This paper improves the traditional Random Walk segmentation algorithm by introducing the general non-locality principle in natural images.The non-local principle indicates that there are Non-adjacent regions with similar representations in the natural image,and these regions can represent each other.To this end,we establish an association representation between the non-neighbor regions in the image;And then construct a non-locally feature incidence Laplacian matrix,combined with the local feature incidence Laplacian matrix constructed in the traditional Random Walk segmentation algorithm.using the local and non-local comprehensive correlation to guide the final The segmentation process.Experiments show that the traditional Random Walk segmentation algorithm is easy to generate pseudo-boundaries in some image regions.The improved algorithm combined with non-local feature association can effectively improve this problem.When segmenting duplicate targets,only individual targets need to be manually labelled,reducing the amount of interaction.(2)Segmentation of similar foreground image sets based on transforming convolutional neural networksThe existing convolutional neural network segmentation algorithm has the following disadvantages:1)It is usually necessary to extract multi-level feature maps of images to complete segmentation,and less attention is paid to geometric correlation between similar foregrounds;2)training of models usually requires a large number of training sets with annotations.Moreover,studies have shown that the prediction of the geometric transformation of the foreground objects can effectively improve the accuracy of object matching and recognition.In view of this,we construct a convolutional neural network segmentation model for predicting foreground pose transformation.Specifically,the input during training is a 4-channel image,and the with the first 3 channels as original images,aiming to extract various features for segmentation.The fourth channel is a transformation mask channel,which is used to estimate the foreground transformations between the target image and the training samples.Transformation masks is obtained by applying multiple affine transformation and thin plate spline deformation to the mask of training samples.When testing,the test image is combined with the mask of the training annotation to predict the foreground mask of the image.The segmentation accuracy statistics and experimental comparisons on the simulation experiments as well as a dataset containing 10 sets of butterfly images show that the proposed method can effectively segment multiple images with similar foregrounds providing a small number of training samples.