Alona Golts

Aviad Aberdam, David Bensaïd, Alona Golts et al. · amazon-science

Reading text in real-world scenarios often requires understanding the context surrounding it, especially when dealing with poor-quality text. However, current scene text recognizers are unaware of the bigger picture as they operate on cropped text images. In this study, we harness the representative capabilities of modern vision-language models, such as CLIP, to provide scene-level information to the crop-based recognizer. We achieve this by fusing a rich representation of the entire image, obtained from the vision-language model, with the recognizer word-level features via a gated cross-attention mechanism. This component gradually shifts to the context-enhanced representation, allowing for stable fine-tuning of a pretrained recognizer. We demonstrate the effectiveness of our model-agnostic framework, CLIPTER (CLIP TExt Recognition), on leading text recognition architectures and achieve state-of-the-art results across multiple benchmarks. Furthermore, our analysis highlights improved robustness to out-of-vocabulary words and enhanced generalization in low-data regimes.

Tsachi Blau, Sharon Fogel, Roi Ronen et al. · amazon-science

The increasing use of transformer-based large language models brings forward the challenge of processing long sequences. In document visual question answering (DocVQA), leading methods focus on the single-page setting, while documents can span hundreds of pages. We present GRAM, a method that seamlessly extends pre-trained single-page models to the multi-page setting, without requiring computationally-heavy pretraining. To do so, we leverage a single-page encoder for local page-level understanding, and enhance it with document-level designated layers and learnable tokens, facilitating the flow of information across pages for global reasoning. To enforce our model to utilize the newly introduced document tokens, we propose a tailored bias adaptation method. For additional computational savings during decoding, we introduce an optional compression stage using our compression-transformer (C-Former),reducing the encoded sequence length, thereby allowing a tradeoff between quality and latency. Extensive experiments showcase GRAM's state-of-the-art performance on the benchmarks for multi-page DocVQA, demonstrating the effectiveness of our approach.

Mor Shpigel Nacson, Aviad Aberdam, Roy Ganz et al. · amazon-science

Vision-Language Models (VLMs) excel in diverse visual tasks but face challenges in document understanding, which requires fine-grained text processing. While typical visual tasks perform well with low-resolution inputs, reading-intensive applications demand high-resolution, resulting in significant computational overhead. Using OCR-extracted text in VLM prompts partially addresses this issue but underperforms compared to full-resolution counterpart, as it lacks the complete visual context needed for optimal performance. We introduce DocVLM, a method that integrates an OCR-based modality into VLMs to enhance document processing while preserving original weights. Our approach employs an OCR encoder to capture textual content and layout, compressing these into a compact set of learned queries incorporated into the VLM. Comprehensive evaluations across leading VLMs show that DocVLM significantly reduces reliance on high-resolution images for document understanding. In limited-token regimes (448$\times$448), DocVLM with 64 learned queries improves DocVQA results from 56.0% to 86.6% when integrated with InternVL2 and from 84.4% to 91.2% with Qwen2-VL. In LLaVA-OneVision, DocVLM achieves improved results while using 80% less image tokens. The reduced token usage allows processing multiple pages effectively, showing impressive zero-shot results on DUDE and state-of-the-art performance on MP-DocVQA, highlighting DocVLM's potential for applications requiring high-performance and efficiency.

Aviad Aberdam, Alona Golts, Michael Elad

Neural networks that are based on unfolding of an iterative solver, such as LISTA (learned iterative soft threshold algorithm), are widely used due to their accelerated performance. Nevertheless, as opposed to non-learned solvers, these networks are trained on a certain dictionary, and therefore they are inapplicable for varying model scenarios. This work introduces an adaptive learned solver, termed Ada-LISTA, which receives pairs of signals and their corresponding dictionaries as inputs, and learns a universal architecture to serve them all. We prove that this scheme is guaranteed to solve sparse coding in linear rate for varying models, including dictionary perturbations and permutations. We also provide an extensive numerical study demonstrating its practical adaptation capabilities. Finally, we deploy Ada-LISTA to natural image inpainting, where the patch-masks vary spatially, thus requiring such an adaptation.

Alona Golts, Daniel Freedman, Michael Elad

Single image dehazing is a critical stage in many modern-day autonomous vision applications. Early prior-based methods often involved a time-consuming minimization of a hand-crafted energy function. Recent learning-based approaches utilize the representational power of deep neural networks (DNNs) to learn the underlying transformation between hazy and clear images. Due to inherent limitations in collecting matching clear and hazy images, these methods resort to training on synthetic data; constructed from indoor images and corresponding depth information. This may result in a possible domain shift when treating outdoor scenes. We propose a completely unsupervised method of training via minimization of the well-known, Dark Channel Prior (DCP) energy function. Instead of feeding the network with synthetic data, we solely use real-world outdoor images and tune the network's parameters by directly minimizing the DCP. Although our "Deep DCP" technique can be regarded as a fast approximator of DCP, it actually improves its results significantly. This suggests an additional regularization obtained via the network and learning process. Experiments show that our method performs on par with large-scale supervised methods.

Alona Golts, Daniel Freedman, Michael Elad

The success of deep learning has been due, in no small part, to the availability of large annotated datasets. Thus, a major bottleneck in current learning pipelines is the time-consuming human annotation of data. In scenarios where such input-output pairs cannot be collected, simulation is often used instead, leading to a domain-shift between synthesized and real-world data. This work offers an unsupervised alternative that relies on the availability of task-specific energy functions, replacing the generic supervised loss. Such energy functions are assumed to lead to the desired label as their minimizer given the input. The proposed approach, termed "Deep Energy", trains a Deep Neural Network (DNN) to approximate this minimization for any chosen input. Once trained, a simple and fast feed-forward computation provides the inferred label. This approach allows us to perform unsupervised training of DNNs with real-world inputs only, and without the need for manually-annotated labels, nor synthetically created data. "Deep Energy" is demonstrated in this paper on three different tasks -- seeded segmentation, image matting and single image dehazing -- exposing its generality and wide applicability. Our experiments show that the solution provided by the network is often much better in quality than the one obtained by a direct minimization of the energy function, suggesting an added regularization property in our scheme.

Alona Golts, Michael Elad

In this paper we present a new approach of incorporating kernels into dictionary learning. The kernel K-SVD algorithm (KKSVD), which has been introduced recently, shows an improvement in classification performance, with relation to its linear counterpart K-SVD. However, this algorithm requires the storage and handling of a very large kernel matrix, which leads to high computational cost, while also limiting its use to setups with small number of training examples. We address these problems by combining two ideas: first we approximate the kernel matrix using a cleverly sampled subset of its columns using the Nyström method; secondly, as we wish to avoid using this matrix altogether, we decompose it by SVD to form new "virtual samples," on which any linear dictionary learning can be employed. Our method, termed "Linearized Kernel Dictionary Learning" (LKDL) can be seamlessly applied as a pre-processing stage on top of any efficient off-the-shelf dictionary learning scheme, effectively "kernelizing" it. We demonstrate the effectiveness of our method on several tasks of both supervised and unsupervised classification and show the efficiency of the proposed scheme, its easy integration and performance boosting properties.