Weak constraint water meter reading identification method

Abstract

The invention relates to a weak-constraint water meter reading recognition method, belongs to the field of artificial intelligence, and provides a feature extraction network based on a convolutional neural network, which has strong feature extraction and expression capabilities. The characteristics of the reading area in the water meter image with a large deviation angle, a large change scale, strong illumination transformation and serious stains can be accurately extracted, and accurate detection of the reading area of the water meter is realized; for a water meter image with an angle deviation absolute value exceeding 90 degrees, the invention provides a direction prediction branch to realize automatic correction and accurate identification of the water meter image with the deviation absolute value exceeding 90 degrees, and the direction prediction branch is particularly effective for the water meter image which is reversed by 180 degrees (inverted). However, the existing water meter identification method is useless for such water meter images.

Description

technical field [0001] The invention belongs to the field of artificial intelligence and relates to a weakly constrained water meter reading recognition method. Background technique [0002] The existing water meter reading recognition methods based on convolutional neural networks are mainly divided into two categories: two-stage water meter recognition methods and end-to-end one-stage water meter recognition methods. The two-stage water meter recognition method usually uses two different neural networks. The first network first detects the reading area in the water meter, and then the second network recognizes the numbers in the detected reading area. The end-to-end one-stage water meter recognition method refers to using only one neural network to directly recognize the numbers in the water meter reading area. However, whether it is a two-stage method or a one-stage method, the existing water meter recognition methods can only better recognize the ideal water meter image...

AI Technical Summary

Problems solved by technology

However, whether it is a two-stage method or a one-stage method, the existing water meter recognition methods can only better recognize the ideal water meter image with a fixed shooting angle or a small change in angle, a small change in illumination, and a relatively clear dial. However, it is helpless for complex and degraded water meter images that appear in actual scenes such as large angle deviation, large illumination changes, and serious dial stains, and the scope of application is very limited.

Images