Research And Implementation Of Intelligent Recognition Of Medical Laboratory Report Image Based On Deep Learning

With the continuous development and upgrading of artificial intelligence,big data technology and computer hardware,more and more people are enjoying the convenience brought by science and technology,and also spawned many emerging industries,such as driver-less,smart city,intelligent medical and so on.At the same time,the exponential growth of various types of data,especially in the medical industry,also puts forward new requirements for digital storage and diagnosis.For quite a long time in the future,relying on the continuous progress of machine learning and deep learning,the interdisciplinary fusion research supported by medical data is still a very important topic in the emerging technology generation and engineering application in our life.In the last two years,the shortage of medical resources and doctors are becoming more and more obvious in the face of the new crown pneumonia.As we all know,the medical examination sheet is very popular in various items of hospital examination,and every Grade-A Tertiary hospital will produce a large number of examination sheet data every day.If all these data are interpreted one by one by working doctors,it will undoubtedly consume the precious time of doctors and patients to a great extent.In order to better assist doctors and patients to understand their own physical condition.In view of the above problems,In this paper,a method based on deep learning is proposed for intelligent detection,recognition and structured output of medical test sheet.Compared with the traditional OCR method,deep learning algorithm design becomes relatively simple,avoiding the complex character feature design problem,and can be well applied to a variety of complex scenes,to recognize Chinese characters,English,numbers,or some special symbols.Through the understanding and learning of the current mainstream image preprocessing,image text detection and recognition,as well as the recognition results of the image layout analysis technology,we found that at present,for users to shoot and upload various images in the natural scene,it is far from reaching the desired recognition results.Because the image shooting in the natural scene is greatly affected by the external environment,it faces many challenges.For example,the image quality is poor,the background is complex,the image contains Chinese characters,English and various characters,etc.Based on the complex and changeable forms of images uploaded by users,in this paper,the traditional adaptive or heuristic algorithm of digital image processing is used to optimize the foreground expression of image and suppress useless noise.Then,on the basis of multi-scale fusion feature pyramid network,a differentiable binary function is added to carry out end-to-end multi-directional text detection.At the same time,the existing advanced deep convolution neural network model is used to carry out transfer learning on different data to realize multiple types of string recognition.During the experiment,the model is continuously fine tuned and some online training model technology is applied to weaken the interference information and inherent angle deviation of some similar characters on the real test sheet.In addition,through model transformation,the models trained by each deep learning library can be combined effectively.Then,for the convenience of later analysis,we need to rearrange recognized text results,filter out useless information for better standardize the effective results of the medical laboratory report image,and finally optimize and effectively deploy the model through reasoning acceleration and multi-threading technology.Finally,the experimental results show that,based on the collected medical dataset,the model can outperform the previous mainstream algorithms in some indicators of detection,location and recognition tasks.