(Geometry Aware) Deep Learning-based Omnidirectional Image Compression

Omnidirectional images are the spherical visual signals that provide a wide, 360◦, view of a scene from a specific position. Such images are becoming increasingly popular in fields like virtual reality and robotics. Compared to conventional 2D images, the storage and badwidth requirements of omnidirectional signals are much higher, due to the specific nature of them. Thus, there is a need for image compression schemes to reduce the dedicated storage space of omnidirectional images. Image compression algorithms can be broadly classified into two groups: lossless and lossy. Lossless schemes are able to reconstruct the exact original data but they cannot reduce the size beyond a specific criteria. Lossy methods are generally better solutions if they do not add a high visual distortion to the reconstructed image, as long as they provide a decent compression rate. If a planar, lossy image compression scheme is applied on omnidirectional images, some problems show up. It is possible to apply a planar compression scheme on a projected version of a 360◦image; however, in these projection schemes (such as equirectangular projection) the sampling rate is different in the poles and the center. Consequently, the filters of the planar compression schemes that do not consider this difference ends in suboptimal result and distortions in the reconstructed images. Recently, with the success of deep neural networks in many image processing tasks, researchers began to use them for the image compression as well. In this study, we propose a deep learning-based method for the compression of omnidirectional images by combining some state of the art approaches from the deep learning-based image compression schemes and some special convolutional layers that take into account the geometry of the omnidirectional image. In comparison to the available methods, it is the first method that can be applied directly on the equirectangularly projected version of omnidirectional images and considers the geometry in the scheme and the layers themselves. To propose this method, different geometry-aware convolutional layers have been tried. We exploited various methods of downsampling and upsampling, such as spherical pooling layers, strided or transposed convolutions, bilinear interpolation, and pixel shuffle. In the end, a method is proposed that benefits from specific spherical convolutional layers which contain sampling methods considering the geometry of omnidirectional images. The sampling positions differ in the different heights of the image based on the nature of the projected omnidirectional image. Additionally, as it benefits from an iterative training method that calculates the residual between the output and input and feeds it again to the network as input of the next iteration, it can provide different compression rates with just one pass of training. Finally, it benefits from a novel method of patching that is well-aligned with the spherical convolution layers and helps the method to run efficiently without a need for a high computational power. The model was compared with a similar architecture without spherical convolutions and spherical patching and showed some improvements. The architecture has been optimized and improved and it has the potential to compete with popular image compression schemes such as JPEG especially in terms of reconstructing the colors.

Graph-based image representation learning

Renata Khasanova

Though deep learning (DL) algorithms are very powerful for image processing tasks, they generally require a lot of data to reach their full potential. Furthermore, there is no straightforward way to impose various properties, given by the prior knowledge about the target task, on the features extracted by a DL model. Therefore, in this thesis we propose several techniques that rely on the power of graph representations to embed prior knowledge inside the learning process. This allows to reduce the solution space and leads to faster optimization convergence and higher accuracy in the representation learning. In our first work, inspired by the ability of a human to correctly classify rotated, shifted or flipped objects, we propose an algorithm that permits to inherently encode invariance to isometric transformations of objects in an image. Our DL architecture is based on graph representations and consists of three novel layers, which we refer to as graph convolutional, dynamic pooling and statistical layers. Our experiments on the image classification tasks show that our network correctly recognizes isometrically transformed objects even though such types of transformation are not seen by the network at training time. Standard DL techniques are typically not able to succeed in solving such a problem without extensive data augmentation. Then, we propose to exploit the properties of graph-based approaches to efficiently process images with various types of projective geometry. In particular, we are interested in increasingly popular omnidirectional cameras, which have a 360 degree field of view. Despite their effectiveness, such cameras create images with specific geometric properties, which require special techniques for efficient processing. We propose an efficient way of adjusting the weights of the graph edges to adapt the filter responses to the geometric image properties introduced by omnidirectional cameras. Our experiments prove that using the proposed graph with properly adjusted edge weights permits to reach better performance as compared to using regular grid graph with equal weights. Finally, the approach described above relies on the isotropic filters, which work well within our transformation invariant architecture for image classification. However, for other problems (e.g. image compression) or even when used without dynamic pooling and statistical layers that are defined within the proposed architecture, these filters are unable to efficiently encode the information about the object. Thus, we introduce a different technique based on anisotropic filters that adapt their shape and size according to the omnidirectional image geometry. The main advantage of this approach compared to the previous one is the ability to encode the orientation of an image pattern, which is important for various tasks such as image compression. Our experiments show that our approach adapts to different image projective geometries and achieves state-of-the-art performance on image classification and compression tasks. Overall we propose several methods, which combine the power of DL and graph signal processing towards incorporating prior information about the target task inside the optimization procedure. We hope that the research efforts presented in this thesis will help the development of efficient DL algorithms that can use various types of prior knowledge to make them efficient even when the available training data is scarce.

EPFL2019

Towards neural network approaches for point cloud compression

Evangelos Alexiou, Touradj Ebrahimi, Kuan Tung

Point cloud imaging has emerged as an efficient and popular solution to represent immersive visual information. However, the large volume of data generated in the acquisition process reveals the need of efficient compression solutions in order to store and transmit such contents. Several standardization committees are in the process of finalizing efficient compression schemes to cope with the large volume of information that point clouds require. At the same time, recent efforts on learning-based compression approaches have been shown to exhibit good performance in the coding of conventional image and video contents. It is currently an open question how learning-based coding performs when applied to point cloud data. In this study, we extend recent efforts on the matter by exploring neural network implementations for separate, or joint compression of geometric and textural information from point cloud contents. Two alternative architectures are presented and compared; that is, a unified model that learns to encode point clouds in a holistic way, allowing fine-tuning for quality preservation per attribute, and a second paradigm consisting of two cascading networks that are trained separately to encode geometry and color, individually. A baseline configuration from the best-performing option is compared to the MPEG anchor, showing better performance for geometry and competitive performance for color encoding at low bit-rates. Moreover, the impact of a series of parameters is examined on the network performance, such as the selection of input block resolution for training and testing, the color space, and the loss functions. Results provide guidelines for future efforts in learning-based point cloud compression.

2020

Directionlets

Vladan Velisavljevic

Efficient representation of geometrical information in images is very important in many image processing areas, including compression, denoising and feature extraction. However, the design of transforms that can capture these geometrical features and represent them with a sparse description is very challenging. Recently, the separable wavelet transform achieved a great success providing a computationally simple tool and allowing for a sparse representation of images. However, in spite of the success, the efficiency of the representation is limited by the spatial isotropy of the wavelet basis functions built in the horizontal and vertical directions as well as the lack of directionality. One-dimensional discontinuities in images (edges and contours), which are very important elements in visual perception, intersect with too many wavelet basis functions leading to a non-sparse representation. To capture efficiently these anisotropic geometrical structures characterized by many more than the horizontal and vertical directions, more flexible multi-directional and anisotropic transforms are required. We present a new lattice-based perfect reconstruction and critically sampled anisotropic multi-directional wavelet transform. The transform retains the separable filtering, subsampling and simplicity of computations and filter design from the standard two-dimensional wavelet transform, unlike in the case of some other existing directional transform constructions (e.g. curvelets, contourlets or edgelets). The corresponding anisotropic basis functions, which we call directionlets, have directional vanishing moments along any two directions with rational slopes. Furthermore, we show that this novel transform provides an efficient tool for non-linear approximation of images, achieving the decay of mean-square error O(N-1.55), which, while slower than the optimal rate O(N-2), is much better than O(N-1) achieved with wavelets, but at similar complexity. Owing to critical sampling, directionlets can easily be applied to image compression since it is possible to use Lagrange optimization as opposed to the case of overcomplete expansions. The compression algorithms based on directionlets outperform the methods based on the standard wavelet transform achieving better numerical results and visual quality of the reconstructed images. Moreover, we have adapted image denoising algorithms to be used in conjunction with an undecimated version of directionlets obtaining results that are competitive with the current state-of-the-art image denoising methods while having lower computational complexity.