A reading recognition method for seven-segment LCD meters based on cycle-consistent adversarial learning
To solve the difficult problem of reading recognition caused by complex illumination,background and other ran-dom interference to meter images in natural scenes,a novel method for seven-segment LCD meter readings based on cy-cle-consistent adversarial learning was presented.M-CRAFT was used as the character detecting network to obtain the read-ing region and the positions of the segmental code characters,and the sub-images of the characters were segmented via post-processing.Then,based on cycle-consistent adversarial domain adaption(CYCADA)network,an improved network for background separation and character recognition,named I-CYCADA,was constructed by considering the perceptual loss and the reconstruction loss as well as designing a classifier.I-CYCADA converted the segmental code character image with com-plex background into a clear and intuitive binary image,which made reading recognition simpler and more accurate.To verify the validity of the proposed method,a dataset composed of images of seven-segment LCD meters in a variety of complex sce-narios was built.Experimental results show that I-CYCADA network can be used in combination with different CNN classifi-cation networks to improve the recognition accuracy of the converted images of segmental code characters.The proposed method achieves a character-level recognition accuracy of 95.39%and a whole reading recognition accuracy of 86.65%on the self-constructed dataset.The recognition effect of difficult samples is effectively improved,and the lightweight design can meet the real-time requirements.
LCD meterreading recognitionadversarial learningdomain adaptationperceptual lossreconstruction loss
万方数据
维普
