Research On Low Contrast Stamping Character Recognition Algorithm Based On Multi-light Illumination

Metal stamping characters has a wide range of applications in industrial automation.It's important to identify metal stamping characters automatically.However,Contrast between low-contrast stamping character and its background is low,and there will be oil stain,fingerprint and rust interference in the long-term use of the process.These factors will increase the difficulty of character segmentation and lead to the failure of character recognition.In view of the above problems,the studies including the following three parts are conducted in this paper:First of all,an image acquisition system based on multi-light illumination was proposed,which captures a group of images in turn by stamping characters in different directions.Based on the analysis of the variation of gray value in the flat and concave areas of the captured images,an image fusion algorithm was proposed to improve the contrast between stamping characters and its background.The difficulty of character segmentation under lower contrast was solved.The oil stain,fingerprints and rust interference were reduced largely.In the experiment,the quality of a single image acquired under the condition of single light source was compared.The result shows that the quality of the image has been greatly improved.Then,the preprocessing and character segmentation were conducted to the fused image.In the preprocessing stage,the fused image was firstly smoothed,binarized and then connected component labeling method was used to eliminate the small disturbing region of the binarized image and the morphological closing operation was used to fill the hollow part of the character strokes.Finally,the tilt angle of character string was detected by the combination of connected domain labeling and line fitting method,and then the character string was rotated correctly.In character segmentation stage,pre-segmentation and fine segmentation of the preprocessed image were conducted to achieve single character by the combination with connected-domain labeling and gray-scale projection,and then each character was normalized to the same size for the next character recognition.Finally,the gray projection histogram and the Histogram of Oriented Gradients(HOG)were used to represent the features of single character.The template matching algorithm and the BP neural network algorithm were respectively used to realize character recognition on the basis of different character features,and the results of different algorithms were compared.Experiments show that the HOG-based BP neural network algorithm is better than the template matching algorithm,and the recognition rate can reach 99.6% for the characters achieved by the algorithm in this paper.Moreover the algorithm is superior to the current research methods of character recognition.