A method for recognizing a table in an image

Abstract

The present application relates to table recognition, and particularly to a table recognition method in an image, which comprises the following steps: pre-processing an input image, and inputting the obtained table image into a neural network; extracting a feature map of the table image by using the neural network, processing the feature map output by the neural network to obtain a target feature map; obtaining a sequence feature of the table based on the target feature map, and obtaining a recognition result of a table category and a cell position according to the sequence feature of the table; and the technical scheme provided by the present application can effectively overcome the defects of low accuracy of table structure and cell position recognition in the prior art.

Claims

1. A method of identifying a table in an image, characterized by: The method comprises the following steps: S1, preprocessing the input image and inputting the obtained table image into a neural network; S2, extracting a feature map of the table image by using the neural network, processing the feature map output by the neural network to obtain a target feature map; S3, obtaining a sequence feature of the table based on the target feature map, and obtaining a recognition result of a table category and a cell position according to the sequence feature of the table; In S3, the sequence feature of the table is obtained based on the target feature map, comprising: transforming the target feature map into a preset size, and adding a position embedding containing feature vector position information to obtain the sequence feature of the table; In S3, the recognition result of the table category and the cell position is obtained according to the sequence feature of the table, comprising: inputting the sequence feature of the table into a transformer module to obtain a table sequence label, performing a softmax operation on the first 37 features in the table sequence label, and the class with the highest score is the table category corresponding to the input image; performing a sigmoid operation on the last 8 data corresponding to each table sequence label to obtain the relative horizontal and vertical coordinates of the four vertices of the cell corresponding to each table sequence label; The table sequence label is simplified according to a simplification rule, and the simplification rule comprises: simplifying a pair of labels in the html code prediction into one table sequence label; eliminating redundant representation in the html code prediction for the merged cell, and only using one table sequence label to represent; not distinguishing whether the cell has content or not.

2. The method of claim 1, wherein: In S1, the input image is preprocessed, comprising: cutting out the table part in the input image, scaling the long side to a preset size while keeping the aspect ratio unchanged, and padding the short side with pixels with a value of 0 to obtain a table image.

3. The method of claim 2, wherein: In S1, the obtained table image is input into the neural network, comprising: inputting the table image into the HRNet_18 neural network in the format of RGB three channels.

4. The method of claim 3, wherein: In S2, the feature map of the table image is extracted by using the neural network, and the target feature map is obtained by processing the feature map output by the neural network, comprising: outputting the feature map of the table image by using the HRNet_18 neural network, adjusting the size of the feature map, stacking the feature maps together, and finally connecting two Conv+BN+Pool layers to obtain the target feature map.

5. The method of claim 4, wherein: The size of the feature map is adjusted by using the HRNet_18 neural network to output the feature map of the table image, comprising: outputting four layers of feature maps feat_Ⅰ, feat_Ⅱ, feat_Ⅲ and feat_Ⅳ of the table image by using the HRNet_18 neural network, and scaling the feature maps feat_Ⅰ, feat_Ⅲ and feat_Ⅳ to the same size as the feature map feat_Ⅱ.

Images