Pointer Meter Recognition Via Image Registration And Visual Saliency Detection

Pointer meter is still widely applied in various fields because of direct and reliable reading, simple structure, low manufacture price, strong anti-electromagnetic interference capability, high precision, easy maintenance, etc. But most of pointer gauges have the nondigital signal output, these gauges could not convert the measuring signal to digital signal. In general, the meter reading can be detected by human. Recording pointer meter reading is a tedious, boring, and highly repetitive work. In some fields containing a large number of pointer gauges, such as power systems, pointer meter verification systems, the accuracy of meter readings largely dependent on responsibility of operator and visual fatigue of operator. It is unavoidable to appear some mistakes and reading errors. If the mistake can be found promptly, the additional work needs to be increased. Otherwise, the mistakes may result in serious consequences. The traditional manual reading way not only causes the waste of human resources, but also may be unable to achieve the desired recognition precision.This paper utilizes computer vision, image processing and pattern recognition technology to realize automatically and accurately pointer type meter readings recognition. This paper bases on Marr vision theory, combines image registration and the visual saliency detection, and concentrates on the key technology in pointer meter recognition, which includes image feature extraction, image registration parameters optimization, saliency region detection, pointer reading recognition model. This paper seeks the breakthroughs and innovations in both reseach methods and ideas, including four aspects as follows:(1) SURF method has a limited performance on images with larger smooth regions. This paper proposes a weighted neighbor-gradient and fuzzy enhancement feature detection approach based on SURF. The basic principle of the presented feature detection method is based on the observation that the gradient feature changes slowly in the smooth region, the extracted feature has poor discriminability in these regions, but feature has strong discriminability in larger gradient change regions. This paper first determines the regions with larger gradient changes, and then extracts image features, which can avoid unstable feature points to effect subsequent image processing procedure. In detail, the proposed method first calculates the image gradient features, and utilizes image gradient features histogram to divide image into multiple tiers, and then constructs a fuzzy function to strengthen different image gradient features in different levels. Secondly, the proposed method exploits the morphological dilation operation and the close operation to detect the regions with the larger gradient changes, then extracts SURF features on these regions. Experimental results illustrate that the proposed approach improves the matched precision on pointer meter images with larger smooth regions. In addition, the presented method is robustness and efficiency on public image databases.(2) Currently, registration accurateness still needs to be further improved. A coarse-to-fine image registration parameter estimation scheme is proposed in this paper. The process of proposed method as follows: the proposed method utilizes the feature point pairs between the reference image and the sensed image to extract the coarse geometrical transform matrixes, and then constructs low-rank model based on coarse geometrical transform matrixes to optimize geometrical transform matrix and estimate the inliers. Finally, the presented method uses these inliers to acquire the optimized transform matrix by an iterative optimization method. Experimental results illustrate that the proposed image registration approach presents a good performance in terms with RRMS and EID, can eliminate geometric distortion of image better.(3) Saliency region detection plays an important role in image pre-processing, but uniformly emphasizing saliency region is still an intractable problem in computer vision. This paper presents a data-driven salient region detection method via multi-feature(included contrast, spatial relationship and background prior, etc.) on absorbing Markov chain, which uses super pixel to extract salient regions, and each super-pixel represents a node. In detail, the proposed saliency detection method first constructs function to calculate absorption probability of each node on absorbing Markov chain. Second the proposed method utilizes image contrast and space relation to model the prior salient map which is provided to foreground salient nodes, and then calculates the saliency of nodes based on absorption probability. Third, the proposed approach also exploits background prior to supply the absorbing nodes and compute the saliency of nodes. Finally, the presented method fuses both the saliency of nodes by cosine similarity measurement method and acquires the ultimate saliency map. This paper tests the proposed saliency detection method on MSRA-B, iCoSeg and SED databases. Experimental results illustrate that the proposed approach presents better robustness and efficiency against the eleven state-of the art algorithms. At the same time, this paper tests the proposed saliency detection method on pointer meter images, our saliency detection method can robustly detects pointer regions.(4) Reading recognition methods of the pointer type meter were vulnerable to illumination variation, and the accuracy was not high. This paper presents a pointer instrument reading method based on visual saliency regions detection. The basic idea of the proposed reading recognition method is that the projections of pointer regions on the vertical axis and on the horizontal axis are different. In detail, the proposed approach first utilizes saliency region detection method to outstand meter pointer region and suppress interference of non-pointer regions. Then the presented method successively rotates meter pointer image, and calculates the maximum value of the projection on the vertical axis in different rotated angles to obtain the rotated angle which rotates the pointer parallel to the horizontal axis. Next, the algorithm rotates meter image to make the pointer paralleling to the horizontal axis, and extracts the top and bottom part pointer region, and determines the maximum projection position on a horizontal axis, and then obtains the pointer deflection angle. Finally, meter reading is derived by using the least squares method to fit a linear function between the pointer deflection angle and the meter scale. This paper tests the presented method on fixed viewpoint camera and pan-tilt camera. Experimental results show that the error between meter readings of the proposed method and manual reading is smaller, and the algorithm is stable and reliable.