Image-Based Electric Consumption Recognition via Multi-Task Learning

This work presents an approach to detect and recognize digits in meter displays by applying the multi-task learning technique. Two convolutional networks are used and divided as tasks to reach different goals. One to detect digits and another to recognize them from different sources, and at the end to return display prediction. In order to validate this methodology, a displays images dataset was labelled. In order to use multi-task learning to reduce the number of failures associated with the reading process. For this, we propose a two-task approach, the first one to detect digits using the networks Faster R-CNN and RetinaNet, and the second stage for recognizing them using Resnet152 network. The initial tests yielded promising results, with 1360 displays, divided into 70% for training and 30% for the test, obtained the following values of Mean Average Precision (mAP), 0.91 (Faster R-CNN) and 0.90 (RetinaNet) on detection and 98.2% accuracy in the classification.