Research On Characters Of Curved Workpiece Recognition Algorithm Based On Deep Learning

As the basic way of information recording,storage and object recognition,character is widely used as the unique identification of workpiece.Intelligent recognition is the basis of realizing production information management.Taking the common metal workpiece as an example,in its service process,the metal surface is vulnerable to oxidation,rust,impact and collision,which will greatly reduce the character recognition rate.Especially for curved surface workpiece,the character text lines on its surface are arranged in a curved shape,the characters on both sides of the workpiece are prone to be distorted,and the reflective phenomenon on the metal surface is serious,which brings great challenges to the positioning and segmentation of character text lines.Therefore,this project takes the low-contrast concave-convex characters on the surface of curved metal workpieces as the research target,and conducts the following research on the key technologies in the imaging detection and recognition of concave-convex characters on the surface of metal workpieces:First,a single-light source imaging image acquisition system is designed,and a circular light source is used directly above the curved metal workpiece with an industrial area scan camera and lens to collect character images.By adjusting the exposure time of the camera,the detail changes of light and dark areas of character surfaces at different near-light positions on surface workpieces are analyzed.A method based on image pyramid fusion algorithm and character detail enhancement is proposed to synthesize high dynamic images from low dynamic image sequences collected with different exposure brightness,which not only improves the contrast between the line of concave and convex characters and the background of the image,but also reduces the specular reflection effect of surface metal workpieces.Then,an curved character text line segmentation detection algorithm based on deep learning is proposed to accurately locate the character image after image enhancement.This text detection network is based on the classical Renet-50 feature extraction network,which fuses feature maps of different scales,obtains the mask of curved character text lines through semantic segmentation,and further obtains the convex hull coordinates of text lines.The experimental results show that the proposed detection algorithm can accurately locate character text lines arranged in different shapes.Finally,a text correction recognition network is proposed to recognize irregular character text lines.The recognition network includes correction module and recognition module,which uses the correction network module to correct curved text lines and partially distorted characters to some extent,and then predicts text character sequences by combining attention-based sequence coding with a learning model of sequence decoding.The experimental results show that the overall recognition rate of concave convex character text lines is as high as 96.8%,which indicates that the algorithm proposed in this paper has strong adaptability and practical value.