Computational Aesthetics and Image Enhancements using Deep Neural Networks

Imaging devices have become ubiquitous in modern life, and many of us capture an increasing number of images every day. When we choose to share or store some of these images, our primary selection criterion is to choose the most visually pleasing ones. Yet, quantifying visual pleasantness is a challenge, as image aesthetics not only correlate with low-level image quality, such as contrast, but also high-level visual processes, like composition and context. For most users, a considerable amount of manual effort and/or professional knowledge is required to get aesthetically pleasing images. Developing automatic solutions thus benefits a large community. This thesis proposes several computational approaches to help users obtain the desired images. The first technique aims at automatically measuring the aesthetics quality, which benefits the users in selecting and ranking images. We form the aesthetics prediction problem as a regression task and train a deep neural network on a large image aesthetics dataset. The unbalanced distribution of aesthetics scores in the training set can result in bias of the trained model towards certain aesthetics levels. Therefore, we propose to add sample weights during training to overcome such bias. Moreover, we build a loss function on the histograms of user labels, thus enabling the network to predict not only the average aesthetics quality but also the difficulty of such predictions. Extensive experiments demonstrate that our model outperforms the previous state-of-the-art by a notable margin. Additionally, we propose an image cropping technique that automatically outputs aesthetically pleasing crops. Given an input image and a certain template, we first extract a sufficient amount of candidate crops. These crops are later ranked according to the scores predicted by the pre-trained aesthetics network, after which the best crop is output to the users. We conduct psychophysical experiments to validate the performance. We further present a keyword-based image color re-rendering algorithm. For this task, the colors in the input image are modified to be visually more appealing according to the keyword specified by users. Our algorithm applies local color re-rendering operations to achieve this goal. A novel weakly-supervised semantic segmentation algorithm is developed to locate the keyword-related regions where the color re-rendering operations are applied. The color re-rendering process benefits from the segmentation network in two aspects. Firstly, we achieve more accurate correlation measurements between keywords and color characteristics, contributing to better re-render rendering results of the colors. Secondly, the artifacts caused by the color re-rendering operations are significantly reduced. To avoid the need of keywords when enhancing image aesthetics, we explore generative adversarial networks (GANs) for automatic image enhancement. GANs are known for directly learning the transformations between images from the training data. To learn the image enhancement operations, we train the GANs on an aesthetics dataset with three different losses combined. The first two are standard generative losses that enforce the generated images to be natural and content-wise similar to the input images. We propose a third aesthetics loss that aims at improving the aesthetics quality of the generated images. Overall, the three losses together direct the GANs to apply appropriate image enhancement operations.

Towards Real-World Super-Resolution using Deep Neural Networks

Ruofan Zhou

Image super-resolution reconstructs a higher-resolution image from the observed low-resolution image. In recent years, machine learning models have been widely employed and deep learning networks have achieved state-of-the-art super-resolution performance. Most of these methods, however, are trained and evaluated on simulated datasets that assume a simple and uniform degradation model. The deep learning models trained on such simulated datasets fail to generalize to practical applications because the actual or real degradation in real-world low-resolution images is much more complex. In this thesis, we propose several approaches to improve the robustness and generalization of deep super-resolution models. The first technique reduces the accumulation of errors in the camera pipeline. We build a deep residual network for learning an end-to-end mapping between raw images and high-resolution images. Our proposed network, trained on high-quality samples, is able to reconstruct in a single step high-quality super-resolved images from low-resolution Bayer mosaics. Extensive experiments show that the proposed method achieves better results than the state-of-the-art techniques, both qualitatively and quantitatively. To resolve the problem of the mismatch between the applied blur kernel in the synthetic dataset and the real-world camera blur, we propose to incorporate blur kernel modeling in the training. We generate the super-resolution training dataset by employing a set of realistic blur kernels estimated from real low-resolution photographs. We build a pool of realistic blur kernels with a generative adversarial network; then, we train a super-resolution network using the low-resolution images constructed with the generated kernels. Our method reconstructs more visually plausible high-resolution images compared to other state-of-the-art methods that rely on a simple degradation model. In order to study the effect of noise on super-resolution, we collect a dataset that contains pairs of noisy low-resolution images and the corresponding high-resolution images by using microscopy. We then benchmark the combinations of denoising methods and super-resolution networks on the dataset. Our experimental results show that the super-resolution networks are sensitive to noise and that the consecutive application of two applications suffers from the accumulation of errors. The benchmark results also suggest that the networks can benefit from joint optimization, hence we use a single network for joint denoising and super-resolution. Our network, trained with a novel texture loss, outperforms any combination of state-of-the-art deep denoising and super-resolution networks. Finally, to take advantage of multi-modal data available in certain applications, we propose a super-resolution system based on the fusion of information from multiple sources. For the application of spectral image super-resolution, we use two downsampled versions of the same image to infer a better high-resolution image for training. We refer to these inputs as a multi-scale modality. As color images are usually taken at a resolution higher than spectral images, we make use of color images as another modality to improve the super-resolution network. We build a pipeline that learns to super-resolve by using multi-scale spectral inputs guided by a color image by combining both modalities. Our proposed method is economical in time and memory consumption yet achieve competitive results.

EPFL2020

Understanding Deep Neural Networks using Adversarial Attacks

Krishna Kanth Nakka

Deep Neural Networks (DNNs) have achieved great success in a wide range of applications, such as image recognition, object detection, and semantic segmentation. Even thoughthe discriminative power of DNNs is nowadays unquestionable, serious concerns have arised ever since DNNs have shown to be vulnerable to adversarial examples craftedby adding imperceptible perturbations to clean images. The implications of these malicious attacks are even more significant for DNNs deployed in real-world systems, e.g.,autonomous driving and biometric authentication. Consequently, an intriguing question that we aim to understand is the underlying behavior of DNNs to adversarial attacks.This thesis contributes to a better understanding of the mechanism of adversarial attacks on DNNs. Our main contributions are broadly in two directions: (1) we proposeinterpretable architectures first to understand the reasons for the success of adversarial attacks and then to improve the robustness of DNNs; (2) we design intuitive adversarialattacks to both mislead and use as a tool to expand our present understanding of DNNs' internal workings and their limitations. In the first direction, we introduce deep architectures that allow humans to interpret the reasoning process of DNNs prediction. Specifically, we incorporate Bag-of-visual-wordsrepresentations from the pre-deep learning era into DNNs using an attention scheme. We find key reasons for adversarial attack success and use these insights to propose anadversarial defense by maximally separating the latent features of discriminative regions while minimizing the contribution of non-discriminative regions in the final prediction.The second direction deals with the design of adversarial attacks to understand DNNs' limitations in a real-world environment. To begin with, we show that existing state-of-the-art semantic segmentation networks that achieve superior performance by exploiting the context are highly susceptible to indirect local attacks. Furthermore, we demonstrate the existence of universal directional perturbations that are quasi-independent of the input template but still successfully fool unknown siamese-based visual object trackers. We then identify that the mid-level filter banks across different backbones bear strong similarities and thus can be potential common ground for attack. We, therefore, learn a generator that disrupts mid-level features with high transferability across different target architectures, datasets, and tasks. In short, our attacks highlight critical vulnerabilities of DNNs, which make their deployment challenging in the real-world environment, even in the extreme case when the attacker is unaware of the target architecture or the targetdata used to train it.Furthermore, we go beyond fooling networks and demonstrate the usefulness of adversarial attacks for studying the internal disentangled representations in self-supervised 3D pose estimation networks. We observe that adversarial manipulation of appearance information in the input image alters the pose output, indicating that the pose code contains appearance information and disentanglement is far from complete. Besides the above contributions, an underlying theme that arises multiple times in this thesis is counteracting the adversarial attacks by detecting them.

EPFL2022

Deep Generative Models and Applications

Tatjana Chavdarova

Over the past few years, there have been fundamental breakthroughs in core problems in machine learning, largely driven by advances in deep neural networks. The amount of annotated data drastically increased and supervised deep discriminative models exceeded human-level performances in certain object detection tasks. The increasing availability in quantity and complexity of unlabelled data also opens up exciting possibilities for the development of unsupervised learning methods. Among the family of unsupervised methods, deep generative models find numerous applications. Moreover, as real-world applications include high dimensional data, the ability of generative models to automatically learn semantically meaningful subspaces makes their advancement an essential step toward developing more efficient algorithms. Generative Adversarial Networks (GANs) are a family of unsupervised generative algorithms that have demonstrated impressive performance for data synthesis and are now used in a wide range of computer vision tasks. Despite this success, they gained a reputation for being difficult to train, which results in a time-consuming and human-involved development process to use them. In the first part of this thesis, we focus on improving the stability and the performances of GANs. Foremost, we consider an alternative training process to the standard one, named SGAN, in which several adversarial âlocalâ pairs of networks are trained independently so that a âglobalâ supervising pair of networks can be trained against them. Experimental results on both toy and real-world problems demonstrate that this approach outperforms standard training in terms of better mitigating mode collapse, stability while converging and that it surprisingly, increases the convergence speed as well. To further reduce the computational footprint while maintaining the stability and performance advantages of SGAN, we focus on training a single pair of adversarial networks using variance reduced gradient. More precisely, we study the effect of the stochastic gradient noise on the training of generative adversarial networks (GANs) and show that it can prevent the convergence of standard game optimization methods, while the batch version converges. We address this issue with two stochastic variance-reduced gradient and extragradient optimization algorithms for GANs, named SVRG-GAN and SVRE, respectively. We observe empirically that SVRE performs similarly to a batch method on the MNIST dataset, while being computationally cheaper, and that SVRE yields more stable GAN training on standard datasets. In the second part of the thesis we present our work on people detection. People detection methods are highly sensitive to occlusions between pedestrians, and using joint visual information from multiple synchronized cameras gives the opportunity to improve detection performance. We address the problem of multi-view people occupancy map estimation using an endâtoâend deep learning algorithm called DeepMCD that jointly utilizes the correlated streams of visual information. DeepMCD empirically outperformed the classical approaches by a large margin. Finally, we present a new large-scale and high-resolution dataset, named WILDTRACK. We provide an accurate joint calibration, as well as a series of benchmark results using baseline algorithms published over the recent months for multi-view detection with deep neural networks, and trajectory estimation using a non-Markovian model.

EPFL2020

Understanding Deep Neural Networks using Adversarial Attacks

Krishna Kanth Nakka

EPFL2022

Deep Generative Models and Applications

Tatjana Chavdarova

EPFL2020

Towards Real-World Super-Resolution using Deep Neural Networks

Ruofan Zhou

EPFL2020