Search

Text recognition is a rapidly evolving task with broad practical applications across multiple industries. However, due to the arbitrary-shape text arrangement, irregular text font, and unintended occlusion of font, this remains a challenging task. To handle images with arbitrary-shape text arrangement and irregular text font, we designed the Discriminative Standard Text Font (DSTF) and the Feature Alignment and Complementary Fusion (FACF). To address the unintended occlusion of font, we propose a Dual Attention Serial Module (DASM), which is integrated between residual modules to enhance the focus on text texture. These components improve text recognition by correcting irregular text and aligning it with the original feature extraction, thus complementing the overall recognition process. Additionally, to enhance the study of text recognition in natural scenes, we developed the VBC Chinese dataset under varying lighting conditions, including strong light, weak light, darkness, and other natural environments. Experimental results show that our method achieves competitive performance on the VBC dataset with an accuracy of 90.8% and an overall average accuracy of 93.8%.

This study focuses on Scene Text Recognition (STR), which plays a crucial role in various applications of artificial intelligence such as image retrieval, office automation, and intelligent transportation systems. Currently, pre-trained vision-language models have become the foundation for various downstream tasks. CLIP exhibits robustness in recognizing both regular (horizontal) and irregular (rotated, curved, blurred, or occluded) text in natural images. As research in scene text recognition requires substantial linguistic knowledge, we introduce the pre-trained vision-language model CLIP and the pre-trained language model Llama. Our approach builds upon CLIP’s image and text encoders, featuring two encoder–decoder branches: one visual branch and one cross-modal branch. The visual branch provides initial predictions based on image features, while the cross-modal branch refines these predictions by addressing the differences between image features and textual semantics. We incorporate the large language model Llama2-7B in the cross-modal branch to assist in correcting erroneous predictions generated by the decoder. To fully leverage the potential of both branches, we employ a dual prediction and refinement decoding scheme during inference, resulting in improved accuracy. Experimental results demonstrate that CLIP-Llama achieves state-of-the-art performance on 11 STR benchmark tests, showcasing its robust capabilities. We firmly believe that CLIP-Llama lays a solid and straightforward foundation for future research in scene text recognition based on vision-language models.

Current research on scene text recognition primarily focuses on languages with abundant linguistic resources, such as English and Chinese. In contrast, there is relatively limited research dedicated to low-resource languages. Advanced methods for scene text recognition often employ Transformer-based architectures. However, the performance of Transformer architectures is suboptimal when dealing with low-resource datasets. This paper proposes a Collaborative Encoding Method for Scene Text Recognition in the low-resource Uyghur language. The encoding framework comprises three main modules: the Filter module, the Dual-Branch Feature Extraction module, and the Dynamic Fusion module. The Filter module, consisting of a series of upsampling and downsampling operations, performs coarse-grained filtering on input images to reduce the impact of scene noise on the model, thereby obtaining more accurate feature information. The Dual-Branch Feature Extraction module adopts a parallel structure combining Transformer encoding and Convolutional Neural Network (CNN) encoding to capture local and global information. The Dynamic Fusion module employs an attention mechanism to dynamically merge the feature information obtained from the Transformer and CNN branches. To address the scarcity of real data for natural scene Uyghur text recognition, this paper conducted two rounds of data augmentation on a dataset of 7267 real images, resulting in 254,345 and 3,052,140 scene images, respectively. This process partially mitigated the issue of insufficient Uyghur language data, making low-resource scene text recognition research feasible. Experimental results demonstrate that the proposed collaborative encoding approach achieves outstanding performance. Compared to baseline methods, our collaborative encoding approach improves accuracy by 14.1%.

Connectionist temporal classification (CTC) is a favored decoder in scene text recognition (STR) for its simplicity and efficiency. However, most CTC-based methods utilize one-dimensional (1D) vector sequences, usually derived from a recurrent neural network (RNN) encoder. This results in the absence of explainable 2D spatial relationship between the predicted characters and corresponding image regions, essential for model explainability. On the other hand, 2D attention-based methods enhance recognition accuracy and offer character location information via cross-attention mechanisms, linking predictions to image regions. However, these methods are more computationally intensive, compared with the 1D CTC-based methods. To achieve both low latency and model explainability via character localization using a 1D CTC decoder, we propose a marginalization-based method that processes 2D feature maps and predicts a sequence of 2D joint probability distributions over the height and class dimensions. Based on the proposed method, we newly introduce an association map that aids in character localization and model prediction explanation. This map parallels the role of a cross-attention map, as seen in computationally-intensive attention-based architectures. With the proposed method, we consider a ViT-CTC STR architecture that uses a 1D CTC decoder and a pretrained vision Transformer (ViT) as a 2D feature extractor. Our ViT-CTC models were trained on synthetic data and fine-tuned on real labeled sets. These models outperform the recent state-of-the-art (SOTA) CTC-based methods on benchmarks in terms of recognition accuracy. Compared with the baseline Transformer-decoder-based models, our ViT-CTC models offer a speed boost up to 12 times regardless of the backbone, with a maximum 3.1% reduction in total word recognition accuracy. In addition, both qualitative and quantitative assessments of character locations estimated from the association map align closely with those from the cross-attention map and ground-truth character-level bounding boxes.

Aiming at the problems of Uyghur oblique deformation, character adhesion and character similarity in scene images, this paper proposes a scene Uyghur recognition model with enhanced visual prediction. First, the content-aware correction network TPS++ is used to perform feature-level correction for skewed text. Then, ABINet is used as the basic recognition network, and the U-Net structure in the vision model is improved to aggregate horizontal features, suppress multiple activation phenomena, better describe the spatial characteristics of character positions, and alleviate the problem of character adhesion. Finally, a visual masking semantic awareness (VMSA) module is added to guide the vision model to consider the language information in the visual space by masking the corresponding visual features on the attention map to obtain more accurate visual prediction. This module can not only alleviate the correction load of the language model, but also distinguish similar characters using the language information. The effectiveness of the improved method is verified by ablation experiments, and the model is compared with common scene text recognition methods and scene Uyghur recognition methods on the self-built scene Uyghur dataset.

Linguistic knowledge helps a lot in scene text recognition by providing semantic information to refine the character sequence. The visual model only focuses on the visual texture of characters without actively learning linguistic information, which leads to poor model recognition rates in some noisy (distorted and blurry, etc.) images. In order to address the aforementioned issues, this study builds upon the most recent findings of the Vision Transformer, and our approach (called Display-Semantic Transformer, or DST for short) constructs a masked language model and a semantic visual interaction module. The model can mine deep semantic information from images to assist scene text recognition and improve the robustness of the model. The semantic visual interaction module can better realize the interaction between semantic information and visual features. In this way, the visual features can be enhanced by the semantic information so that the model can achieve a better recognition effect. The experimental results show that our model improves the average recognition accuracy on six benchmark test sets by nearly 2% compared to the baseline. Our model retains the benefits of having a small number of parameters and allows for fast inference speed. Additionally, it attains a more optimal balance between accuracy and speed.

Scene text recognition (STR) has been a hot research field in computer vision, aiming to recognize text in natural scenes using computers. Currently, attention-based encoder–decoder frameworks struggle to precisely align feature regions with the target object when dealing with complex and low-quality images, a phenomenon known as attention drift. Additionally, with the rise of Transformer, the increasing size of parameters results in higher computational costs. In order to solve the above problems, based on the latest research results of Vision Transformer (ViT), we utilize an additional position-enhancement branch to alleviate attention drift and dynamically fused position information with visual information to achieve better recognition accuracy. The experimental results demonstrate that our model achieves a 3% higher average recognition accuracy on the test set compared to the baseline. Meanwhile, our model maintains the advantage of a small number of parameters and fast inference speed, achieving a good balance between accuracy, speed, and computational load.

The real-life scene images exhibit a range of variations in text appearances, including complex shapes, variations in sizes, and fancy font properties. Consequently, text recognition from scene images remains a challenging problem in computer vision research. We present a scene text recognition methodology by designing a novel feature-enhanced convolutional recurrent neural network architecture. Our work addresses scene text recognition as well as sequence-to-sequence modeling, where a novel deep encoder–decoder network is proposed. The encoder in the proposed network is designed around a hierarchy of convolutional blocks enabled with spatial attention blocks, followed by bidirectional long short-term memory layers. In contrast to existing methods for scene text recognition, which incorporate temporal attention on the decoder side of the entire architecture, our convolutional architecture incorporates novel spatial attention design to guide feature extraction onto textual details in scene text images. The experiments and analysis demonstrate that our approach learns robust text-specific feature sequences for input images, as the convolution architecture designed for feature extraction is tuned to capture a broader spatial text context. With extensive experiments on ICDAR2013, ICDAR2015, IIIT5K and SVT datasets, the paper demonstrates an improvement over many important state-of-the-art methods.

Reading Indian scene texts is complex due to the use of regional vocabulary, multiple fonts/scripts, and text size. This work investigates the significant differences in Indian and Latin Scene Text Recognition (STR) systems. Recent STR works rely on synthetic generators that involve diverse fonts to ensure robust reading solutions. We present utilizing additional non-Unicode fonts with generally employed Unicode fonts to cover font diversity in such synthesizers for Indian languages. We also perform experiments on transfer learning among six different Indian languages. Our transfer learning experiments on synthetic images with common backgrounds provide an exciting insight that Indian scripts can benefit from each other than from the extensive English datasets. Our evaluations for the real settings help us achieve significant improvements over previous methods on four Indian languages from standard datasets like IIIT-ILST, MLT-17, and the new dataset (we release) containing 440 scene images with 500 Gujarati and 2535 Tamil words. Further enriching the synthetic dataset with non-Unicode fonts and multiple augmentations helps us achieve a remarkable Word Recognition Rate gain of over 33% on the IIIT-ILST Hindi dataset. We also present the results of lexicon-based transcription approaches for all six languages.

Text recognition in natural scene images has always been a hot topic in the field of document-image related visual sensors. The previous literature mostly solved the problem of horizontal text recognition, but the text in the natural scene is usually inclined and irregular, and there are many unsolved problems. For this reason, we propose a scene text recognition algorithm based on a text position correction (TPC) module and an encoder-decoder network (EDN) module. Firstly, the slanted text is modified into horizontal text through the TPC module, and then the content of horizontal text is accurately identified through the EDN module. Experiments on the standard data set show that the algorithm can recognize many kinds of irregular text and get better results. Ablation studies show that the proposed two network modules can enhance the accuracy of irregular scene text recognition.

This paper presents a comprehensive survey on Arabic cursive scene text recognition. The recent years’ publications in this field have witnessed the interest shift of document image analysis researchers from recognition of optical characters to recognition of characters appearing in natural images. Scene text recognition is a challenging problem due to the text having variations in font styles, size, alignment, orientation, reflection, illumination change, blurriness and complex background. Among cursive scripts, Arabic scene text recognition is contemplated as a more challenging problem due to joined writing, same character variations, a large number of ligatures, the number of baselines, etc. Surveys on the Latin and Chinese script-based scene text recognition system can be found, but the Arabic like scene text recognition problem is yet to be addressed in detail. In this manuscript, a description is provided to highlight some of the latest techniques presented for text classification. The presented techniques following a deep learning architecture are equally suitable for the development of Arabic cursive scene text recognition systems. The issues pertaining to text localization and feature extraction are also presented. Moreover, this article emphasizes the importance of having benchmark cursive scene text dataset. Based on the discussion, future directions are outlined, some of which may provide insight about cursive scene text to researchers.