Vision transformers (ViTs) have pushed the state-of-the-art for various visual recognition tasks by patch-wise image tokenization followed by stacked self-attention operations. Employing self-attention modules results in a quadratic complexity in both computation and memory usage. Various attempts on approximating the self-attention computation with linear complexity have thus been made in Natural Language Processing. However, an in-depth analysis in this work reveals that they are either theoretically flawed or empirically ineffective for visual recognition. We identify that their limitations are rooted in retaining the softmax self-attention during approximations. Specifically, conventional self-attention is computed by normalizing the scaled dot-product between token feature vectors. Preserving the softmax operation challenges any subsequent linearization efforts. Under this insight, a SOftmax-Free Transformer (abbreviated as SOFT) is proposed for the first time. To eliminate the softmax operator in self-attention, a Gaussian kernel function is adopted to replace the dot-product similarity. This enables a full self-attention matrix to be approximated via a low-rank matrix decomposition. The robustness of our approximation is achieved by calculating its Moore-Penrose inverse using a Newton-Raphson method. Further, an efficient symmetric normalization is introduced on the low-rank self-attention for enhancing model generalizability and transferability. Extensive experiments on ImageNet, COCO and ADE20K show that our SOFT significantly improves the computational efficiency of existing ViT variants. Crucially, with a linear complexity, much longer token sequences are permitted in SOFT, resulting in superior trade-off between accuracy and complexity.
Heterogeneity exists in most camera images. This heterogeneity manifests itself across the image space as varied Moire ringing, motion-blur, color-bleaching or lens based projection distortions. Moreover, combinations of these image artifacts can be present in small or large pixel neighborhoods, within an acquired image. Current camera image processing pipelines, including deep trained versions, tend to rectify the issue applying a single filter that is homogeneously applied to the entire image. This is also particularly true when an encoder-decoder type deep architecture is trained for the task. In this paper, we present a structured deep learning model that solves the heterogeneous image artifact filtering problem. We call our deep trained model the Patch Subspace Variational Autoencoder (PS-VAE) for Camera ISP. PS-VAE does not necessarily assume uniform image distortion levels nor similar artifact types within the image. Rather, our model attempts to learn to cluster different patches extracted from images into artifact type and distortion levels, within multiple latent subspaces (e.g. Moire ringing artifacts are often a higher dimensional latent distortion than a Gaussian motion blur artifact). Each image's patches are encoded into soft-clusters in their appropriate latent sub-space, using a prior mixture model. The decoders of the PS-VAE are also trained in an unsupervised manner for each of the image patches in each soft-cluster. Our experimental results demonstrates the flexibility and performance that one can achieve through improved heterogeneous filtering. We compare our results to a conventional one-encoder-one-decoder architecture.
When the trained physician interprets medical images, they understand the clinical importance of visual features. By applying cognitive attention, they apply greater focus onto clinically relevant regions while disregarding unnecessary features. The use of computer vision to automate the classification of medical images is widely studied. However, the standard convolutional neural network (CNN) does not necessarily employ subconscious feature relevancy evaluation techniques similar to the trained medical specialist and evaluates features more generally. Self-attention mechanisms enable CNNs to focus more on semantically important regions or aggregated relevant context with long-range dependencies. By using attention, medical image analysis systems can potentially become more robust by focusing on more important clinical feature regions. In this paper, we provide a comprehensive comparison of various state-of-the-art self-attention mechanisms across multiple medical image analysis tasks. Through both quantitative and qualitative evaluations along with a clinical user-centric survey study, we aim to provide a deeper understanding of the effects of self-attention in medical computer vision tasks.
Overparameterized neural networks enjoy great representation power on complex data, and more importantly yield sufficiently smooth output, which is crucial to their generalization and robustness. Most existing function approximation theories suggest that with sufficiently many parameters, neural networks can well approximate certain classes of functions in terms of the function value. The neural network themselves, however, can be highly nonsmooth. To bridge this gap, we take convolutional residual networks (ConvResNets) as an example, and prove that large ConvResNets can not only approximate a target function in terms of function value, but also exhibit sufficient first-order smoothness. Moreover, we extend our theory to approximating functions supported on a low-dimensional manifold. Our theory partially justifies the benefits of using deep and wide networks in practice. Numerical experiments on adversarial robust image classification are provided to support our theory.
ImUnity is an original deep-learning model designed for efficient and flexible MR image harmonization. A VAE-GAN network, coupled with a confusion module and an optional biological preservation module, uses multiple 2D-slices taken from different anatomical locations in each subject of the training database, as well as image contrast transformations for its self-supervised training. It eventually generates 'corrected' MR images that can be used for various multi-center population studies. Using 3 open source databases (ABIDE, OASIS and SRPBS), which contain MR images from multiple acquisition scanner types or vendors and a large range of subjects ages, we show that ImUnity: (1) outperforms state-of-the-art methods in terms of quality of images generated using traveling subjects; (2) removes sites or scanner biases while improving patients classification; (3) harmonizes data coming from new sites or scanners without the need for an additional fine-tuning and (4) allows the selection of multiple MR reconstructed images according to the desired applications. Tested here on T1-weighted images, ImUnity could be used to harmonize other types of medical images.
In this paper, we introduce FairFaceGAN, a fairness-aware facial Image-to-Image translation model, mitigating the problem of unwanted translation in protected attributes (e.g., gender, age, race) during facial attributes editing. Unlike existing models, FairFaceGAN learns fair representations with two separate latents - one related to the target attributes to translate, and the other unrelated to them. This strategy enables FairFaceGAN to separate the information about protected attributes and that of target attributes. It also prevents unwanted translation in protected attributes while target attributes editing. To evaluate the degree of fairness, we perform two types of experiments on CelebA dataset. First, we compare the fairness-aware classification performances when augmenting data by existing image translation methods and FairFaceGAN respectively. Moreover, we propose a new fairness metric, namely Frechet Protected Attribute Distance (FPAD), which measures how well protected attributes are preserved. Experimental results demonstrate that FairFaceGAN shows consistent improvements in terms of fairness over the existing image translation models. Further, we also evaluate image translation performances, where FairFaceGAN shows competitive results, compared to those of existing methods.
Video prediction models often combine three components: an encoder from pixel space to a small latent space, a latent space prediction model, and a generative model back to pixel space. However, the large and unpredictable pixel space makes training such models difficult, requiring many training examples. We argue that finding a predictive latent variable and using it to evaluate the consistency of a future image enables data-efficient predictions because it precludes the necessity of a generative model training. To demonstrate it, we created sequence completion intelligence tests in which the task is to identify a predictably changing feature in a sequence of images and use this prediction to select the subsequent image. We show that a one-dimensional Markov Contrastive Predictive Coding (M-CPC_1D) model solves these tests efficiently, with only five examples. Finally, we demonstrate the usefulness of M-CPC_1D in solving two tasks without prior training: anomaly detection and stochastic movement video prediction.
Existing computer vision research in artwork struggles with artwork's fine-grained attributes recognition and lack of curated annotated datasets due to their costly creation. To the best of our knowledge, we are one of the first methods to use CLIP (Contrastive Language-Image Pre-Training) to train a neural network on a variety of artwork images and text descriptions pairs. CLIP is able to learn directly from free-form art descriptions, or, if available, curated fine-grained labels. Model's zero-shot capability allows predicting accurate natural language description for a given image, without directly optimizing for the task. Our approach aims to solve 2 challenges: instance retrieval and fine-grained artwork attribute recognition. We use the iMet Dataset, which we consider the largest annotated artwork dataset. In this benchmark we achieved competitive results using only self-supervision.
Curvilinear structure segmentation plays an important role in many applications. The standard formulation of segmentation as pixel-wise classification often fails to capture these structures due to the small size and low contrast. Some works introduce prior topological information to address this problem with the cost of expensive computations and the need for extra labels. Moreover, prior work primarily focuses on avoiding false splits by encouraging the connection of small gaps. Less attention has been given to avoiding missed splits, namely the incorrect inference of structures that are not visible in the image. In this paper, we present DTU-Net, a dual-decoder and topology-aware deep neural network consisting of two sequential light-weight U-Nets, namely a texture net, and a topology net. The texture net makes a coarse prediction using image texture information. The topology net learns topological information from the coarse prediction by employing a triplet loss trained to recognize false and missed splits, and provides a topology-aware separation of the foreground and background. The separation is further utilized to correct the coarse prediction. We conducted experiments on a challenging multi-class ultrasound scan segmentation dataset and an open dataset for road extraction. Results show that our model achieves state-of-the-art results in both segmentation accuracy and continuity. Compared to existing methods, our model corrects both false positive and false negative examples more effectively with no need for prior knowledge.
Photoacoustic imaging (PAI) can image high-resolution structures of clinical interest such as vascularity in cancerous tumor monitoring. When imaging human subjects, geometric restrictions force limited-view data retrieval causing imaging artifacts. Iterative physical model based approaches reduce artifacts but require prohibitively time consuming PDE solves. Machine learning (ML) has accelerated PAI by combining physical models and learned networks. However, the depth and overall power of ML methods is limited by memory intensive training. We propose using invertible neural networks (INNs) to alleviate memory pressure. We demonstrate INNs can image 3D photoacoustic volumes in the setting of limited-view, noisy, and subsampled data. The frugal constant memory usage of INNs enables us to train an arbitrary depth of learned layers on a consumer GPU with 16GB RAM.