Research On Real-time Positioning Technology Of Carousel Archives Based On Fisheye Lens Image Recognition

Basically,the high-density storage of paper files adopts the method of revolving inventory.Checking whether the file is in the library and whether the file storage location is correct is a tedious and important task in daily management.The traditional manual inventory method is labor-intensive and error-prone.This theme uses the current more mature image recognition technology to research the technology of digital label of the file box image recognition,which has very important significance in reducing the cost of file management and improving the accuracy of the location and recognition of the inventory.The market application of the research results has a good prospect.Due to the narrow internal space of the rotary file cabinet,and mostly use embedded systems as the core of data acquisition and control.This theme proposes a real-time disk library solution based on the collection and recognition technology of fisheye lens digital image tags.When the digital label recognition and positioning is implemented in this solution,the image blur caused by the fisheye lens distortion and the large time cost of deep learning are two difficulties to be solved.This theme first uses fisheye lens to collect digital image tags,and completes the image distortion correction in less time through the checkerboard calibration.Due to the phenomenon of uneven distribution of light and dark and blurred handwriting in the edge area after the correction,the theme researches and designs the optimal image binarization solution to obtain a clearer binarized digital label image.After using the contour tracking algorithm to perform label segmentation and single-digit character segmentation on digital label images,the paper uses the Le Net-5 network to identify single-character images.In order to improve the operation speed of single-character image recognition,the paper improves the Le Net-5 network to reduce the operation parameter variables;in order to overcome the blurring of the handwriting and the binarization of the strokes caused by binarization,the accuracy of the phenomenon is reduced,the paper has carried out on the training data set of missing characters.The enhanced processing improves the recognition accuracy rate.In addition,the paper also studies the target detection algorithm using YOLOv3-tiny network to directly complete the digital label recognition of the grayscale image of the fisheye lens after correction.Finally,single character recognition and YOLOv3-tiny direct target detection and recognition constitute double complementary recognition of digital image labels,which further improves the accuracy of digital image labels.The main work and innovations of the thesis are as follows:(1)The optimal value method of window scale and noise floor in adaptive binarization is studied,which solves the image caused by the uneven distribution of light and dark of the fisheye lens image after correction,and the blurred writing in the edge area,such as quality degradation.(2)The convolution kernel of the original Le Net-5 network is improved,from the original single-layer 5 x 5 convolution kernel to a two-layer 3 x 3 convolution kernel,which reduces the amount of model parameters and improves the real-time performance of the algorithm.(3)Aiming at the decrease of the accuracy caused by the incomplete binarized strokes caused by the blurred writing,the paper randomly cuts the digital characters and enhances the training data set,so that this model has a better generalization ability for the recognition of the digital characters lost in some pixels.(4)Optimize the original YOLOv3-tiny network structure,reduce the downsampling rate of the network,and improve the accuracy of the model's detection of small targets(file digital character labels).