A multitask cascading convolutional neural network for high-accuracy pointer meter automatic recognition in outdoor environments. (14th February 2023)

• Record Type: Journal Article
• Authors: Liu, Fang; Pan, Lei; Gao, Rui; Zhang, Liyang; Pang, Yi; Ning, Xucheng; Zhang, Hao; Liu, Kunlei
• Is Part Of: Measurement science & technology. Volume 34:Number 5(2023)
• Publication Date: 2023-02-14
• Subjects: convolutional neural network (CNN) -- meter detection -- meter reading -- YOLOV4; Physical measurements -- Periodicals; Scientific apparatus and instruments -- Periodicals; Equipment and Supplies -- Periodicals; Science -- instrumentation -- Periodicals; Technology -- instrumentation -- Periodicals; Mesures physiques -- Périodiques; Physical measurements; Scientific apparatus and instruments; Periodicals; 502.87
• DOI: 10.1088/1361-6501/acb003
• Languages: English
• ISSNs: 0957-0233

Abstract

Abstract: Pointer meter automatic recognition (PMAR) in outdoor environments is a challenging task. Due to variable weather and uneven lighting factors, hand-crafted features or shallow learning techniques have low accuracy in meter recognition. In this paper, a multitask cascading convolutional neural network (MC-CNN) is proposed to improve the accuracy of meter recognition in outdoor environments. The proposed MC-CNN uses cascaded CNN, including three stages of meter detection, meter cropping and meter reading. Firstly, the YOLOV4 Network is used for meter detection to quickly determine the meter location from captured images. In order to accurately cluster pointer meter prior boxes in the YOLOV4 Network, an improved K-means algorithm is presented to further enhance the detection accuracy. Then, the detected meter images are cropped out of the captured images to remove redundant backgrounds. Finally, a meter-reading network based on an adaptive attention residual module (AARM) is proposed for reading meters from cropped images. The proposed AARM not only contains an attention mechanism to focus on essential information and efficiently diminish useless information, but also extracts information features from meter images adaptively. The experimental results show that the proposed MC-CNN can effectively achieve outdoor meter recognition, with high recognition accuracy and low relative error. The recognition accuracy can reach 92.6%. The average relative error is 2.5655%, which is about 3% less than the error in other methods. What is more, the proposed approach can obtain rich information about the type, limits, units and readings of the pointer meter and can be used when multiple pointer meters exist in one captured image simultaneously. Additionally, the proposed approach can significantly improve the accuracy of the recognized readings, and is also robust to natural environments.