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.

We propose a framework for sequence-to-sequence contrastive learning (SeqCLR) of visual representations, which we apply to text recognition. To account for the sequence-to-sequence structure, each feature map is divided into different instances over which the contrastive loss is computed. This operation enables us to contrast in a sub-word level, where from each image we extract several positive pairs and multiple negative examples. To yield effective visual representations for text recognition, we further suggest novel augmentation heuristics, different encoder architectures and custom projection heads. Experiments on hand-written text and on scene text show that when a text decoder is trained on the learned representations, our method out-performs non-sequential contrastive methods. In addition, when the amount of supervision is reduced, SeqCLR significantly improves performance compared with supervised training, and when fine-tuned with 100% of the labels, our method achieves state-of-the-art results on standard hand-written text recognition benchmarks.

Scene Text Recognition (STR), the task of recognizing text against complex image backgrounds, is an active area of research. Current state-of-the-art (SOTA) methods still struggle to recognize text written in arbitrary shapes. In this paper, we introduce a novel architecture for STR, named Selective Context ATtentional Text Recognizer (SCATTER). SCATTER utilizes a stacked block architecture with intermediate supervision during training, that paves the way to successfully train a deep BiLSTM encoder, thus improving the encoding of contextual dependencies. Decoding is done using a two-step 1D attention mechanism. The first attention step re-weights visual features from a CNN backbone together with contextual features computed by a BiLSTM layer. The second attention step, similar to previous papers, treats the features as a sequence and attends to the intra-sequence relationships. Experiments show that the proposed approach surpasses SOTA performance on irregular text recognition benchmarks by 3.7% on average.

The iterative linear expansion of threshold framework, or iLET, offers a new approach for solving image restoration problems under sparsity assumptions. Instead of estimating the reconstructed image directly, the iLET paradigm parametrizes the reconstruction process as a linear combination of elementary thresholding functions and optimizes over their coefficients. Here, we rely on the fast and accurate convergence of iLET, and propose an extension of this framework, under the assumption that the reconstructed object is approximately piece-wise constant. This assumption leads to a new total-variation framework of iLET. We demonstrate the applicability of our technique to bio-medical imaging problems, such as computerized tomography reconstruction. Our technique surpasses state-of-the-art reconstructions in terms of PSNR and SSIM, while offering an automatic way for tuning its regularization parameter.

We present a composite suite of technologies for spectral coexistence of existing communication and radar systems using the Xampling framework. For a stand-alone communication system, we consider a cognitive radio (CRo) that receives multiband signals with unknown carrier frequencies and directions of arrival, and demonstrate joint spectrum sensing via CompreSsed CArrier and Direction-ofarrival Estimation (CaSCADE) with an L-shaped configuration of two uniform linear arrays. For radars operating in bands with widespread spectral interference, we present an X-band prototype of cognitive sub-Nyquist multiple input multiple output (MIMO) radar (SUMMeR). The prototype allows sampling in both spatial and spectral domains at sub-Nyquist rates and cognitively transmits over multiple narrow subbands. Finally, we demonstrate Spectral Coexistence via Xampling (SpeCX) technology that shows joint operation of both - cognitive radio and cognitive monostatic radar - over a common spectrum. Our solutions to individual and joint operation of communication and radar systems supersede existing spectrum sharing technologies that require a compromise over performance of one of the systems.

In the context of Cognitive Radio (CR), opportunistic transmissions can exploit temporarily vacant spectral bands. Efficient and reliable spectrum sensing is key in the CR process. CR receivers traditionally deal with wideband signals with high Nyquist rates and low Signal to Noise Ratios (SNRs). Sub-Nyquist sampling of such signals has been proposed for efficient sampling in CRs. The modulated wideband converter (MWC) is an example of such a sampling scheme. It is composed of an analog front-end, that aliases the signal intentionally before sampling it at a low rate. The signal can then be digitally reconstructed from the low rate samples, using the known relation between the samples and the original signal. Unfortunately, in real hardware implementation, this relation becomes unknown. Physical effects have a considerable impact on the sampling process, and as a consequence, the signal cannot be reliably recovered. In this paper, we present an efficient automated calibration algorithm that builds the actual transfer function of the system, without any prior knowledge. We then present a new, MWC based, CR prototype, on which the calibration algorithm was tested. Experiments on our hardware prototype, based on an embedded proprietary card, show that our calibrated transfer function leads to signal reconstruction whereas the theoretical one fails. Our specification complies with CR requirements of the IEEE standard 802.22 and was experimentally verified with different modulations. It vastly improves a previous prototype in terms of bandwidth, higher maximal frequency and coping with lower SNR.