Character Recognition of Low Illumination Product Marking Based on Vision

Abstract

This paper is mainly aimed at recognition of the horizontal indeterminate-length marking characters of products under low illumination conditions. We first obtain images to be processed from local galleries or the camera. Then, we apply multi-scale Retinex algorithm in HSV space to improve the brightness of images and denoising technology to improve the quality of images. Finally, we realize text detection through CTPN network model and character recognition through CRNN network model. End-to-end character recognition is achieved. The experimental results show that the accuracy of the CTPN network model is 80.91%, and the accuracy of the CRNN network model is 61.79% respectively.