Research On Text Detection And Recognition Of Infusion Labels Based On Deep Learning

In recent years,text detection and text recognition in medical scenarios have gradually become one of the research hotspots in the field of artificial intelligence.At present,most pharmacies have been relying on manual completion of dispensing,dispensing,replenishment and inventory management,which is easy to cause medical problems due to staff fatigue,visual deviation and other human factors.Therefore,the intelligent text detection and recognition system with high efficiency and high precision has very important application value.This thesis takes infusion labels of medical drugs(infusion bag labels and infusion bottle labels)as the research object,aiming at the two phenomena of uneven illumination and curved text in practical application scenarios,studies the following four aspects: uneven illumination processing,text detection,correction and recognition of curved text and text detection and recognition system of infusion labels:(1)Mainly aiming at the phenomenon of local bright spots in label text images caused by uneven illumination in infusion bags under natural conditions,an adaptive enhancement algorithm based on Retinex was proposed to adjust the brightness of areas with high exposure in images.Firstly,multi-scale Retinex is used to extract the light component in the brightness(V)component image.Then,the adaptive correction function proposed in this thesis is used to realize the adaptive brightness correction of the component.Finally,the corresponding component is fused and the final enhanced image is obtained by converting HSV to RGB.Experiments show that the improved algorithm can effectively deal with the luminance problem of the overexposed area in the image,and realize the uniform correction of image luminance,which lays a foundation for the subsequent text recognition.(2)Aiming at the problems of Text bending and skewing of infusion labels,an adaptive text area representation arbitrary shape(ATRR)text detection method based on improved text candidate network(text-RPN)was proposed.In this method,the backbone network SE-VGG16 in Text-RPN is replaced by the SE-Res Net50 network formed by the SENet attention module in the Res Net50 residual branch,and the Text detection network is built on this basis.The experimental results show that the improved text detection network improves the detection speed by about 33% and the detection accuracy by a certain extent.(3)A CRNN text recognition method based on the improved TPS-STN correction network was proposed,aiming at the text bending phenomenon of infusion bottle labels.Combined with the number of boundary points of the adaptive bounding text boundary box obtained during the detection of curved text in(2)above,the method improves the input parameters of the positioning network in the spatial transformation network(STN)through the proposed adaptive setting method,so that the number of reference points can be set more reasonably according to the shape of text in the infusion label image.Thus,adaptive bending correction is realized.The experiment shows that the improved TPS-STN network has a good correction effect.When the corrected image is input into the CRNN text recognition network for recognition,the corrected text recognition accuracy is increased by 5.07% compared with before correction.(4)Built the hardware and software platform of infusion label text detection and recognition system,transplanted the system into Raspberry PI,and completed the GUI interface design with Py Qt5.The experimental results show that the system has good detection and recognition results and has certain practical application value.