Text understanding from natural scene images has progressively gained much interest in computer vision due to the frequent emergence of handheld or wearable devices. Sometimes, these devices capture low-quality images. As a result, the image can be disrupted with unpredictable blur caused by text movement and camera shake. Further, the problem is integrated with edge intelligence, an emerging paradigm that drives the computing services, applications, and tasks from mainstream cloud to network edge. The purpose is to minimize the bandwidth cost and latency and improves privacy. Hence, this research introduces a novel semi pipelined technique to address such challenging issue for edge intelligence. The fundamental contributions of this work are as given: (i) after text image enhancement, the natural text images are blurred to create a synthetic dataset; (ii) we introduced image-based 2D Radix-4 DIT FFT and the inverse to deblur the blurred images; (iii) after that, text understanding process is applied on the recovered images. Firstly, the text localization and segmentation are taken out using a novel open contour and 4-connected edge-based region approach. Secondly, the recovered images are classified into text and non-text classes adopting multimodal feature representation. Thirdly, Character Labeling Convolution Neural Network (CL-CNN) model is introduced for character labeling by extracting deep features to work fine on discriminative and ambiguous text. Finally, the experiments validated that the proposed framework achieved promising results on ICDAR 2003, SVT, and IIIT5K compared with standard techniques substantially and from blur images efficiently and flexibly.

由于手持或可穿戴设备的频繁出现，自然场景图像中的文本理解逐渐引起了计算机视觉的兴趣。有时，这些设备会捕获低质量的图像。因此，图像可能会因文本移动和相机抖动导致的不可预测的模糊而中断。此外，该问题与边缘智能相结合，边缘智能是一种新兴范式，可驱动计算服务、应用程序和任务从主流云到网络边缘。目的是最小化带宽成本和延迟并提高隐私。因此，本研究引入了一种新颖的半流水线技术来解决边缘智能的这一具有挑战性的问题。这项工作的基本贡献如下：（i）在文本图像增强之后，对自然文本图像进行模糊处理以创建合成数据集；(ii) 我们引入了基于图像的 2D Radix-4 DIT FFT 和逆去模糊图像；(iii) 之后，对恢复的图像应用文本理解过程。首先，使用一种新颖的开放轮廓和基于 4 个连接边缘的区域方法进行文本定位和分割。其次，恢复的图像采用多模态特征表示分为文本类和非文本类。第三，引入了字符标签卷积神经网络（CL-CNN）模型，通过提取深层特征来进行字符标签，以在区分性和歧义文本上很好地工作。最后，实验验证了所提出的框架在 ICDAR 2003、SVT、