Research On Real-time Chip Character Recognition Using Deep Learning

Chip character is the unique identification(ID)of chip,and chip character recognition system plays a vital role in all kinds of chip facilities,such as rearrangement facility.Currently,most algorithms for chip character recognition applied in industry are still focusing on hand-crafted features,which are not accurate,effective and robust.Although current state-of-art methods for text reading based on deep learning can automatically extract text features,they have several pipelines which might lead to serious error propagation and difficult training process.Instead,a chip character recognition system based on deep learning is proposed which learns to recognize chip characters directly from full images involving no complex pre-process and post-process.Our approach is simple,easy to train and test with high precision,speed and robustness.Firstly,an end-to end recognizer for small character is proposed.It is difficult to deal with chip characters at extremely small size.Thus a method based on multi-scale features is introduced by mixing feature maps of higher layers and lower layers.The proper architecture,default boxes,pyramid feature network and loss functions are carefully designed according to specific chip character.Then a benchmark dataset and evaluation protocols are introduced for chip character recognition.Secondly,blind deblurring of chip image is raised.Unavoidable motion blur in chip images may cause serious failures in character recognition.A network based on CGAN is trained to learn the mapping between sharp and blurred images with well-designed architecture and loss function.A benchmark dataset is also introduced by randomly generating trajectories and evaluation protocols for image deblurring.Thirdly,an approach for data augmentation of chip character recognition is proposed.Due to deficiency and poor diversity of chip dataset,it is tough to learn to recognize “character minority”.Character label is taken as the constraint condition of the proposed network based on CGAN,which can not only generate character images,but also can control the category of generated character.It is convenient to generate label files and enlarge chip character dataset.All in all,the overall chip character recognition system is combined of data augmentation,deblurring system and chip character recognizer.Experimental results show that the proposed system achieves the lowest inaccuracy of 0.33%,the lowest missing rate of 0.09% and the lowest running time of 19.9ms per image,reaching high demands for real-time chip character recognition in industrial application.