Research On Image Key Information Extraction Algorithm Based On Computer Vision

“A picture is worth a thousand words”,this proverb emphasizes the richness of information contained in an image.Images are playing an increasingly important role because of their rich information.In order to achieve various purposes,people need to extract key information conveyed from images.In recent years,due to the rapid development of deep learning and its close ties with computer vision,it becomes possible to realize the automation of image key information extraction.This paper studies two image key information extraction algorithms.First,the algorithm for extracting water level information is studied.In order to achieve the trade-off between algorithm stability,complexity,generalization,accuracy,etc.,three water level information extraction methods,method of differential edge detection,classification method based on dictionary learning and classification method based on deep learning,are proposed,and they are carried out through experiments to compare their performance.Method of differential edge detection firstly differentiates images of adjacent frames,then performs threshold segmentation on the product of pixel mean value and standard deviation in the row to find the position of water level.Complexity of this method is low and average error is small,but it has poor stability for its strict requirements of image light.Classification method based on dictionary learning and classification method based on deep learning share the same idea,converting the problem of water level information extraction to a problem of image classification.The difference between the two is that classification method based on dictionary learning uses dictionary learning for image classification,and the other uses deep learning for image classification.Compared with method of differential edge detection,classification methods add a step of image training,so they have higher algorithm complexity,but average error of classification methods are smaller and the methods are more stable.Finally,comparing the two classification methods,classification method based on deep learning has higher computational complexity,but it generalizes better for different scenes.Second,the algorithm for extracting code information is studied.A two-stage code information extraction algorithm based on object detection is proposed.The first stage uses object detection to locate code blocks,and the second stage uses object detection and image classification to locate and identify code characters in the code blocks.Experiments on the test set show that the first stage of the algorithm has fewer missed detections of code blocks and locates code blocks accurately,while the second stage of the algorithm has fewer missed detections of characters and recognites characters accurately.The comprehensive results of the two-stage experiments on the test set show that the accuracy of code information extraction reaches more than 95%.Finally,the image classification network in water level information extraction algorithm and the code block locating network in code information extraction algorithm are lightened.The standard convolution operation is replaced by a depth-wise separable convolution.Deep-wise separable convolution has similar functions with standard convolution,but requires less calculation and parameters.Experiments show that as to both water level information extraction and code information extraction,there is no significant drop in performance when the amount of network calculations and parameters are greatly reduced.