Abstract: Braille is a tool for visually impaired individuals to learn knowledge and skills. Normal individuals often have little knowledge of braille, resulting in numerous communication barriers between normal and visually impaired individuals. A natural scene braille segment image dataset was first constructed, which includes 1 157 braille segment images with different aspect ratios and backgrounds, as well as corresponding label information; Subsequently, the characteristics of braille in natural scene images were analyzed, and a natural scene braille recognition method was proposed based on the SSD framework. The proposed method selects feature layers for braille recognition, designs CNN structures, default box sizes, braille character labels, image input strategies and loss functions to improve the precision of braille recognition based on the small size and fixed aspect ratio of braille characters. Moreover, a pixel level attention mechanism is designed based on the characteristic that braille points are located in the edge region of braille characters, to improve the recall of braille recognition. Experimental results show the proposed method achieves a value of 0.903 in H and a braille character recognition speed of 66.22 frame/s. Compared with SSD, Faster R-CNN, EAST, and CNN based methods named EAST-Edge and UNet-Braille on the constructed natural scene braille segment image datasets, the braille recognition performance of the proposed method has been significantly improved.