http://www.ipol.im/feed/ IPOL Preprints — Latest public preprints from IPOL. Natalia Bottaioli, Daniel Parres, Yung-Hsin Chen http://www.ipol.im/pub/pre/587/ http://www.ipol.im/pub/pre/587/ Thu, 23 Apr 2026 17:03:27 +0200 2026-04-23T15:03:27Z This article focuses on analyzing several aspects of the handwritten text recognition (HTR) models belonging to the TrOCR family introduced by Minghao Li et al. in [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models, AAAI Conference on Artificial Intelligence, 2023]. The TrOCR models are designed to recognize single lines of English text using a transformer-based encoder-decoder architecture. All models incorporate a pre-trained vision transformer as the encoder and a pre-trained text transformer as the decoder. The encoder is responsible for extracting key features from the image, while the decoder autoregressively transcribes the text, subword by subword, based on the extracted features. The authors report state-of-the-art performance across different text types, including handwritten, scene, and printed text. Our analysis has several objectives. The first one is to gain a better understanding of the training process and the data used for producing the handwritten models. The second one is to highlight and explore the functionality and limitations of the TrOCR model in the context of HTR, which poses unique challenges such as variations in individual writing styles. Additionally, we propose an architecture diagram that helped us better understand what the model actually does with the input text line image, which we hope will be useful for the research community using TrOCR. **This is an MLBriefs article, the source code has not been reviewed!**<br> **The original source code is available [[here|https://huggingface.co/microsoft/trocr-base-handwritten]] (last checked 2026/04/23).**<br> Yi Liu, Qi Xie, Yu Guo, Guoqing Chen, Boying Wu, Deyu Meng, Jean-Michel Morel, Qiyu Jin, Michael Kwok-Po Ng http://www.ipol.im/pub/pre/607/ http://www.ipol.im/pub/pre/607/ Tue, 04 Nov 2025 14:38:41 +0100 2025-11-04T13:38:41Z Color image denoising is a critical task in computer vision, often hindered by the underutilization of inter-channel correlations, resulting in color distortion and loss of fine details. We propose QMSANet, a Quaternion Multi-Scale Attention Network, to address these challenges by leveraging quaternion operations for color image denoising. Operating in the quaternion domain, QMSANet preserves channel dependencies across all processing stages, enhancing noise suppression and detail retention. The network comprises three innovative modules: the Quaternion Multi-Scale Sparse Block (QMSB) for extracting multi-scale features with sparsity enforcement, the Quaternion Stacked Enhancement Block (QSEB) for refining deep features through inter-channel interactions, and the Lightweight Quaternion Attention Block (LQAB) for adaptively focusing on salient features with minimal computational overhead. These modules collectively mitigate color deviation, detail loss, and edge artifacts. Extensive experiments on benchmark datasets demonstrate that QMSANet outperforms state-of-the-art denoising models in both synthetic and real-world noisy conditions. Typically, a blind denoiser exhibits diminished performance in comparison to a non-blind denoiser. However, QMASNet-B, a blind denoiser constructed based on our model, also surpasses most of the comparison models. At sigma = 15, QMASNet and QMASNet-B achieve PSNR improvements of 0.53 dB and 0.50 dB, respectively, compared to the state-of-the-art method on CBSD68. Visual comparisons further highlight its ability to preserve structural details. QMSANet offers a balanced, efficient solution for high-quality color image denoising, with significant potential for real-world applications. Marina Gardella http://www.ipol.im/pub/pre/558/ http://www.ipol.im/pub/pre/558/ Wed, 17 Jul 2024 15:19:29 +0200 2024-10-05T22:16:56Z The camera trace is a signal embedded in the image during the image formation process which implicitly encodes information about the camera processing chain. In the article 'Camera Trace Erasing', Chen et al. propose a Siamese Trace Erasing method aiming at extracting those traces. In order to do so, the authors design a novel hybrid loss for network training. This hybrid loss is defined as a combination of three different losses: the embedded similarity loss, the truncated fidelity loss and the cross-identity loss. In this article we briefly explore the method and its results. **This is an MLBriefs article, the source code has not been reviewed!**<br> **The original source code is [[available here|https://github.com/ngchc/CameraTE]] (last checked 2024/10/06).**<br> Gabriele Simone, Mauro Fiorentini, Alessandro Rizzi http://www.ipol.im/pub/pre/531/ http://www.ipol.im/pub/pre/531/ Sun, 31 Mar 2024 20:13:30 +0200 2025-10-02T10:48:15Z The original presentation of Retinex, a spatial color correction and image enhancement algorithm modeling the Human Vision System, uses paths to explore the image in search of a local reference white point. Here we present a spatial color algorithm, called Termite Retinex, with an alternative way to explore local properties of Retinex, replacing random paths with a colony of agents, which uses swarm intelligence to explore the image, determining in this way the locality of its filtering. We show the efficacy of Termite Retinex for unsupervised image enhancement, dynamic range stretching and color correction for digital images. Mahdi Ranjbar, Aissa Abdelaziz, Mohammad Ali Jauhar http://www.ipol.im/pub/pre/532/ http://www.ipol.im/pub/pre/532/ Wed, 13 Mar 2024 19:46:46 +0100 2024-03-13T18:46:46Z Natural image matting refers to the process of estimating the foreground opacity, also known as the alpha matte, of an input image and extracting the foreground layer. It is a fundamental and challenging task in computer vision and finds applications in various fields related to image processing such as film and image editing, advertising, and medical diagnosis. Numerous approaches have been proposed to address this challenge incorporating both traditional and deep learning techniques. In this paper, we reproduce the paper "A Closed Form Solution to Natural Image Matting", also referred to as "Closed Form Matting". The proposed algorithm is based on local smoothness assumptions on foreground and background colors, the color-line model, and provides a simple closed-form solution that requires sparse scribble constraints instead of more cumbersome and hard-to-develop trimap constraints. We present extended derivations, a Python-based implementation, an online demo, along with comparison on contemporary methods and ablation study on the effects of hyper-parameters. Perceval Beja-Battais, Gaëtan Serré, Sophia Chirrane http://www.ipol.im/pub/pre/469/ http://www.ipol.im/pub/pre/469/ Mon, 05 Jun 2023 20:33:05 +0200 2023-06-05T18:33:05Z In this paper, we define an experimental context in which we tested the performances of LIPO and AdaLIPO, two global optimization algorithms for Lipschitz functions, introduced in [C. Malherbe and N. Vayatis, Global optimization of lipschitz functions, 2017]. We provide experimental proofs of the efficiency of those algorithms, led numerical statistical analysis of our results, and suggested two intuitive improvements from the vanilla version of the algorithms, referred as LIPO-E and AdaLIPO-E. Within our test bench, these improvements allow the algorithms to converge significantly faster and whenever they struggle to find a better maximizer. Finally, we defined the scope of application of LIPO and AdaLIPO. We show that they are very prone to the curse of dimensionality and tend quickly to Pure Random Search when the dimension increases. We provide source code for LIPO, AdaLIPO, and our enhanced versions Zhe Zheng, Gabriele Facciolo, Pablo Arias http://www.ipol.im/pub/pre/464/ http://www.ipol.im/pub/pre/464/ Tue, 28 Feb 2023 19:44:12 +0100 2023-03-14T12:42:14Z Recently, the field of image and video denoising has undergone a revolution thanks to deep learning approaches. These methods outperform traditional model-based approaches in almost every image/video restoration problem. In this paper, we propose an implementation of a recent approach proposed for video denoising, namely Efficient Multi-stage Video Denoising method (EMVD). The method has a lightweight and interpretable architecture consisting of three stages: temporal fusion, denoising, and refinement stages. We reproduce this method and propose three modifications aimed at improving its performance. (1) We apply motion compensation to make better use of temporal redundancy, (2) we apply variance stabilization to help this lightweight network deal with signal-dependent noise and (3) we decouple occlusion detection and fusion weights prediction. We evaluate the original method and the proposed modifications on a task of raw video denoising. Arnaud Barral http://www.ipol.im/pub/pre/436/ http://www.ipol.im/pub/pre/436/ Fri, 25 Nov 2022 09:21:14 +0100 2022-11-25T08:21:14Z Temporal high pass filter methods are a family of methods for Fixed Pattern Noise (FPN) reduction. They are recursive real time methods that apply a high-pass temporal filter to remove the FPN. FPN is a temporally coherent noise present on video due to the non-uniformity response of the sensors. It is a common problem for infrared videos and can degrade the quality of the observation. In this work we will study and compare three classical temporal high pass filter FPNR methods.