Deep Learning
Z. Raisi; V. M. Nazarzehi Had; E. Sarani; R. Damani
Abstract
Background and Objectives: Research on right-to-left scripts, particularly Persian text recognition in wild images, is limited due to lacking a comprehensive benchmark dataset. Applying state-of-the-art (SOTA) techniques on existing Latin or multilingual datasets often results in poor recognition performance ...
Read More
Background and Objectives: Research on right-to-left scripts, particularly Persian text recognition in wild images, is limited due to lacking a comprehensive benchmark dataset. Applying state-of-the-art (SOTA) techniques on existing Latin or multilingual datasets often results in poor recognition performance for Persian scripts. This study aims to bridge this gap by introducing a comprehensive dataset for Persian text recognition and evaluating SOTA models on it.Methods: We propose a Farsi (Persian) text recognition (FATR) dataset, which includes challenging images captured in various indoor and outdoor environments. Additionally, we introduce FATR-Synth, the largest synthetic Persian text dataset, containing over 200,000 cropped word images designed for pre-training scene text recognition models. We evaluate five SOTA deep learning-based scene text recognition models using standard word recognition accuracy (WRA) metrics on the proposed datasets. We compare the performance of these recent architectures qualitatively on challenging sample images of the FATR dataset.Results: Our experiments demonstrate that SOTA recognition models' performance declines significantly when tested on the FATR dataset. However, when trained on synthetic and real-world Persian text datasets, these models demonstrate improved performance on Persian scripts.Conclusion: Introducing the FATR dataset enhances the resources available for Persian text recognition, improving model performance. The proposed datasets, trained models, and code is available at https://github.com/zobeirraisi/FATDR.
Computer Vision
Z. Raisi; J. Zelek
Abstract
Background and Objectives: Signage is everywhere, and a robot should be able to take advantage of signs to help it localize (including Visual Place Recognition (VPR)) and map. Robust text detection & recognition in the wild is challenging due to pose, irregular text instances, illumination variations, ...
Read More
Background and Objectives: Signage is everywhere, and a robot should be able to take advantage of signs to help it localize (including Visual Place Recognition (VPR)) and map. Robust text detection & recognition in the wild is challenging due to pose, irregular text instances, illumination variations, viewpoint changes, and occlusion factors.Methods: This paper proposes an end-to-end scene text spotting model that simultaneously outputs the text string and bounding boxes. The proposed model leverages a pre-trained Vision Transformer based (ViT) architecture combined with a multi-task transformer-based text detector more suitable for the VPR task. Our central contribution is introducing an end-to-end scene text spotting framework to adequately capture the irregular and occluded text regions in different challenging places. We first equip the ViT backbone using a masked autoencoder (MAE) to capture partially occluded characters to address the occlusion problem. Then, we use a multi-task prediction head for the proposed model to handle arbitrary shapes of text instances with polygon bounding boxes.Results: The evaluation of the proposed architecture's performance for VPR involved conducting several experiments on the challenging Self-Collected Text Place (SCTP) benchmark dataset. The well-known evaluation metric, Precision-Recall, was employed to measure the performance of the proposed pipeline. The final model achieved the following performances, Recall = 0.93 and Precision = 0.8, upon testing on this benchmark.Conclusion: The initial experimental results show that the proposed model outperforms the state-of-the-art (SOTA) methods in comparison to the SCTP dataset, which confirms the robustness of the proposed end-to-end scene text detection and recognition model.
