Text Detection And Recognition In Invoice-style Data

Bills and form data play an important role in daily life and are important carriers of business operation and financial management.Efficient and accurate recognition of bill form data is an important part of data analysis and electronization.Recently,the mainstream deep learning technology has excellent effect and application in natural scene text detection and recognition.However,this kind of technology can't process the bill form data well,and the difficulties lie in: 1.Different from the text detection of natural scenes,bill form detection is essentially a multiclassification detection task.The meaning of each field needs to be recognized to extract the required information.2.The density of the fields is very high and the sizes vary greatly.3.Although different forms have different templates,there are numerous types.In the detection task,we introduced the idea of extremum point detection to solve the problem of high-density field and scale difference of ticket form sample data.For multi-classification detection,we introduce category voting mechanism;In addition,we propose a migration training strategy,which can complete the model migration training with only a small amount of target area training data.In the end,our detection model achieves the state of the art in the horizontal comparison of non-category detection,and the model has strong generalization ability,which can still maintain reliable detection effect in the face of complex samples.In the recognition task,we integrate the attention mechanism into the sliding window model to segment and recognize the single word in the sequence in a semi-supervised way.Finally,through horizontal comparison,our model achieved the State of the Art in both serial-level accuracy and single-word accuracy,and was nearly 1.5 times as fast as the CRNN-CTC series model using the same backbone network,and 1.2 times faster than the attention-based model.