2023 | Book

# Image Co-segmentation

Authors: Avik Hati, Rajbabu Velmurugan, Sayan Banerjee, Subhasis Chaudhuri

Publisher: Springer Nature Singapore

Book Series : Studies in Computational Intelligence

2023 | Book

Authors: Avik Hati, Rajbabu Velmurugan, Sayan Banerjee, Subhasis Chaudhuri

Publisher: Springer Nature Singapore

Book Series : Studies in Computational Intelligence

This book presents and analyzes methods to perform image co-segmentation. In this book, the authors describe efficient solutions to this problem ensuring robustness and accuracy, and provide theoretical analysis for the same. Six different methods for image co-segmentation are presented. These methods use concepts from statistical mode detection, subgraph matching, latent class graph, region growing, graph CNN, conditional encoder–decoder network, meta-learning, conditional variational encoder–decoder, and attention mechanisms. The authors have included several block diagrams and illustrative examples for the ease of readers. This book is a highly useful resource to researchers and academicians not only in the specific area of image co-segmentation but also in related areas of image processing, graph neural networks, statistical learning, and few-shot learning.

Advertisement

Abstract

If one is given several images sourced from various places and under different contexts, but having at least ‘something’ in common, then the extraction of this common object in all these images is known as image co-segmentation. It is the problem of segmenting objects with similar features from more than one image. In this monograph, we discuss efficient solutions to this problem ensuring robustness and high accuracy and provide theoretical analysis for the same.

Abstract

This chapter provides a review of the literature related to image co-segmentation and datasets used for evaluation.

Abstract

This chapter describes some concepts that will be instrumental in developing the co-segmentation algorithms of this monograph. They are superpixel segmentation, label propagation, maximum common subgraph computation, convolution neural networks, variational inference and few-shot learning.

Abstract

This chapter describes an unsupervised and computationally efficient image co-segmentation algorithm for an image pair. The method is based on subgraph matching. First, the images are represented as region adjacency graphs through image superpixels. Then the maximum common subgraph (MCS) is computed, and the corresponding subgraphs in the respective images give the co-segmented objects. Next, this algorithm is extended to co-segment more than two images containing a common object by performing co-segmentation of every unique image pair followed by co-segmentation of initial results in a hierarchical manner.

Abstract

This chapter describes a robust framework to solve image co-segmentation where the common object is not present in all the images in the set. The co-segmentation problem for N images considers the very general setting that only an unknown M (\(\le N\)) number of images contain the co-segmentable common object(s). This problem, in general, requires solving \(\mathcal {O}(N 2^{N-1})\) MCS matching steps. A computationally efficient method is described in this chapter that requires only \(\mathcal {O}(N)\) matching steps. This is achieved by performing maximally occurring common subgraph matching (MOCS) of the images in the set. The first step is to obtain a coarse co-segmentation of images using superpixel clustering that gives the common object partially. Then a ‘latent class graph’ (LCG) is constructed by combining the graphical representations of the partial object in all constituent images. Next the LCG is used for performing region growing on the graphs of individual images to obtain the common object completely. The co-segmentation method requires only \(\mathcal {O}(N)\) image matching operations, instead of \(\mathcal {O}(N 2^{N-1})\), and yet ensures globally consistent matching across images.

Abstract

This chapter describes a robust solution for the multi-image co-segmentation problem under a classification problem setup, with the classes being the common foreground and the remaining regions (backgrounds) in the image set, but yet in an unsupervised framework. First a method to find the dominant mode in the high dimensional feature space of image superpixels is explained. The superpixels having features in close proximity to the computed mode are initially labeled as the common foreground class, while superpixels having a high probability of belonging to the background are assigned the labels of multiple background classes. Then a measure for the separation between the common foreground regions and the background regions in the feature space is defined. Using the labeled samples, a discriminative space is obtained that maximizes the separation between the classes. In this discriminative space, samples from the same class come closer and those from different classes go far apart. Next, label propagation is performed in this discriminative space to assign labels to the unlabeled regions as well as update the labels of the already labeled regions. This discriminative space computation and label propagation are iterated till convergence, finally yielding the common foreground object.

Abstract

This chapter describes a supervised method for solving the co-segmentation problem for an image pair by training a graph convolutional neural network (GCNN)-based classifier. First each image pair is represented as a weighted graph through superpixel segmentation, and intra-image and inter-image superpixel feature similarities. The network takes the graph as input, and learns to classify each node of the graph into either the common foreground or the background. During training, an additional classifier is used that learns to classify the entire graph into the semantic class the common object belongs to. The entire network is trained end-to-end using a combination of the binary cross-entropy loss and the categorical cross-entropy loss.

Abstract

This chapter describes a deep convolutional neural network-based co-segmentation model through an end-to-end training of a conditional siamese encoder-decoder network. The network is composed of a pair of VGG-16-based encoder-decoder network with shared weights, a metric learning network and a decision network. The metric learning network finds an optimum latent feature space where objects of the same class are closer and that of different classes are separated by a certain margin. Depending on the extracted features, the decision network predicts the presence or absence of common objects in the input image pair, and the encoder-decoder network produces co-segmentation masks accordingly. The model is completely class agnostic and does not require any semantic information.

Abstract

This chapter describes a few-shot learning approach to tackle the small sample size problem encountered in learning co-segmentation models with small datasets like iCoseg and MSRC. The multi-image co-segmentation algorithm explained in this chapter involves a class agnostic meta-learning strategy by generalizing the model to new classes given only a small number of training samples for each new class. This is achieved through a directed variational inference cross-encoder, which learns a continuous embedding space to ensure better similarity learning.