Invoice text detection method

Abstract

The invention belongs to the technical field of text recognition, and particularly relates to an invoice text detection method, which comprises the following steps: S1, designing an invoice key pointskeleton detection model, detecting key points of an invoice textbox and establishing a connection relationship; S2, solving a coordinate label of each key point according to the original labeling information of the invoice for training; S3, training an invoice key point skeleton detection model; S4, inputting an image, and detecting key points of the invoice full image by using the trained invoice key point skeleton detection model; S5, solving vertex coordinates of the corresponding textbox according to the coordinate tags of the key points and the relationship between the coordinate tags; S6, cutting the textbox and outputting the textbox according to the vertex coordinates of the textbox, the positions of the key points covering all character positions which may be detected, the position parameters of the textbox can be directly deduced, and the processing speed is high; and no matter which direction the invoice image deforms, the numerical value of the key point can be automatically adjusted according to the image deformation, so that the robustness is high.

Description

technical field [0001] The invention belongs to the technical field of text recognition, in particular to an invoice text detection method. Background technique [0002] The text detection of invoices is a key environment in the task of invoice recognition. How to detect the required text lines on the bill surface quickly and preparedly is the key to the bill recognition technology. [0003] One of the existing methods is to use traditional machine learning methods, first use image processing technology to denoise, grayscale, and binarize the input bill image, and then use morphology, contour extraction, etc. to roughly determine the position of the bill Or give the candidate area, and finally use the method of local feature combined with classifier to screen the candidate area to determine the final text area. However, these methods have low accuracy and poor robustness, and are not suitable for commercial use. [0004] The second existing method is to use deep learning t...

AI Technical Summary

Problems solved by technology

However, these methods have low accuracy and poor robustness, and are not suitable for commercial use.

[0004] The second existing method is to use deep learning technology to detect the text box of the invoice, such as using a general detection model to detect the text of the bill. Thanks to the powerful feature extraction ability of the convolutional neural network, the effect of this method is better than that of the traditional method. It is greatly improved, but the general detection model can only detect horizontal rectangular boxes. Since the bill text boxes are often inclined, the output detection boxes often leave a relatively large gap, which affects the subsequent character classification effect.

If the direction correction module is added after the detection module, it will increase the redundancy of the system and affect the running speed

[0005] The third existing method is to first detect the key points of the invoice to obtain two positions with obvious characteristics in the invoice, and then use these two positions as a basis to correct the direction of the invoice and crop the irrelevant area to obtain the invoice image after cutting and alignment. Then detect the text box on this picture, generally using a two-step method, first correct the position of the invoice through key point detection and related processing operations, and then output all text boxes through text detection on the corrected image, the method has good robustness , good accuracy, but the overall process is redundant, and there is still room for optimization in terms of process and speed

Images