Research On Low Contrast 3D Character Visual Recognition Algorithm In Complex Background

The contrast between 3D stamping characters and the background is very low,and in industrial environments,it is subject to some complicated interferences such as paint,oil,dust,rust,scale,etc.,and the traditional single image acquisition method is inevitable.In order to realize the automatic detection and recognition of 3D stamping characters,this paper has carried out the following research:Firstly,a multi-directional illumination system is designed,which sequentially lights up the light source to collect an image in different directions.It can enhance the feature information of the character edge by using the concave and convex characteristics of the 3D stamping characters.Using the acquired four images containing different feature details,a method based on Laplacian pyramid image fusion is used for image fusion.The multi-scale image fusion method can avoid the loss of details of interest features.This can restore the scene appearance more realistically and reliably.Experimental comparison shows that compared with the traditional image acquisition effect,the proposed method can restore and retain the characteristics of 3D stamping characters more.This method can improve the contrast and discrimination between characters and backgroundThen,a character detection algorithm based on deep learning with improved post-processing output is proposed to detect the 3D stamping characters of the fused image.The network uses a classic VGG-16 architecture for feature extraction.The post-processing output has six prediction maps,including text sequence prediction map,single-character prediction map,and four numerical prediction maps.Four numerical prediction maps were used to calculate weighted average for four vertex coordinates of the text area.Experiments show that the proposed improved detection algorithm has a precision of 99.0%for 3D stamping characters and a recall rate of 96.1%,and has obvious advantages over other text detection algorithms for 3D stamping characters.Finally,the convolutional recurrent neural network(CRNN)is used to recognize the sequence characters.The network mainly includes a convolution feature extraction module,a recurrent neural network feature sequence coding module and a CTC decoding module.The performance of the whole algorithm is tested under 10,000 samples.The overall precision of detection is 99.02%,and the overall recognition rate is 98.80%,which indicates that the proposed algorithm for 3D stamping characters has good practical value.