There are many different recognition methods for English characters, and there are a lot of books mentioned those recognition methods, but font always a main factor that affecting recognizability. The recognition process starts from document’s layout analysis and pre-processing to text segmentation, optical character recognition (OCR) and post-processing, these stages both will affect recognition result. However, most stages are hard to influence by font except OCR stage, and there are more than 50 common English fonts, so this cause the current OCR can not fit every font, this is similar with most classic binarization methods cannot face all situations, cannot separate all images from the foreground and background that the background is cleared, and get the foreground we want. Moreover, how do humans learn how to write the English alphabet? It starts with strokes. Thus, what are strokes? It’s a combination of lines in different directions, so those different direction lines are an important feature. The traditional Artificial Intelligence (AI) Network will be trained by the pre-processed graphics to get the desired model; however, when the traditional methods face different fonts that might because different fonts do not have clear features reduce network learning efficiency, and also will affect the accuracy of subsequent predictions. To efficiently enhance the learning process, this research adopts the stroke feature and is further trained by a lightweight AI network to accelerate predictions , referred to as VLNet (Very Lightweight Network). Strokes feature extraction can reduce the amount of graphics data, which means that the input of training is reduced and the architecture of the network is simplified. Those factors can make the network smaller and more efficient, and also reduce the computing requirements, so it will suitable for edge computing. The experimental result shown that the accuracy of unknown fonts is 97.81% is better than LeNet’s 96.67% and AlexNet’s 91.15%, and the accuracy of know fonts is 99.1% is better than LeNet’s 98.59% and AlexNet’s 95.48%. Then when using Raspberry Pi 4B per character predicted time is between 4.44ms and 4.61ms.