Deep Learning for Image-based Automatic Dial Meter Reading: Dataset and Baselines

Towards Image-Based Automatic Meter Reading in Unconstrained Scenarios: A Robust and Efficient Approach

“…We also focus on studies related to digit-based meters, even though there are some recent works that addressed the recognition of dial meters [13], [14], [23]. Such works usually explore the angle between the pointer and the dial to perform the reading.…”

“…With the advances of deep learning-based techniques and the availability of ever larger datasets, in many cases it is time-consuming to divide the datasets multiple times and then average the results among multiple runs. Hence, public datasets introduced in recent years commonly have a single division of the images into training, validation and test sets [3], [14]. In the same direction, we randomly split the Copel-AMR dataset as follows: 5,000 images for training, 5,000 images for testing and 2,500 images for validation, following the split protocol (i.e., 40%/40%/20%) used in the UFPR-AMR dataset.…”

“…In other words, our AMR system can tolerate less accurate detections of the counter region, as the corners will be later located and the counter rectified. Taking this into account as well as the importance of having an efficient system in real-world applications, we initially decided to use the Fast-YOLOv4 model [36] for this task since, despite being much smaller than state-of-the-art object detectors, Fast-YOLO variants (also known as Tiny-YOLO) are still able to detect some objects quite precisely [37] and have been employed in various research areas in recent years [14], [38], [39].…”

“…Content may change prior to final publication. [14]. Nevertheless, such a low threshold (IoU > 0.5) was deliberately defined by Everingham et al [58] to account for inaccuracies in bounding boxes in the training data, as defining the bounding box for a highly nonconvex object (e.g., a person with arms and legs spread) is somewhat subjective.…”

“…The idea behind image-based AMR, which is an specific scenario of scene text detection and recognition, is that the aforementioned inspection can be carried out automatically, reducing mistakes introduced by the human factor and saving manpower [3], [13]. As pointed out by Salomon et al [14], the consumers themselves can capture photos of meters using a mobile device (e.g., a cell phone or a tablet). This eliminates the need for employees of the service company traveling around to perform local meter reading at each consumer unit, resulting in cost savings (especially in rural areas).…”