Abstract
This paper is concerned with the problem of short circuit detection in infrared image for metal electrorefining with an improved Faster Region-based Convolutional Neural Network (Faster R-CNN). To address the problem of insufficient label data, a framework for automatically generating labeled infrared images is proposed. After discussing factors that affect sample diversity, background, object shape, and gray scale distribution are established as three key variables for synthesis. Raw infrared images without fault are used as backgrounds. By simulating the other two key variables on the background, different classes of objects are synthesized. To improve the detection rate of small scale targets, an attention module is introduced in the network to fuse the semantic segment results of U-Net and the synthetic dataset. In this way, the Faster R-CNN can obtain rich representation ability about small scale object on the infrared images. Strategies of parameter tuning and transfer learning are also applied to improve the detection precision. The detection system trains on only synthetic dataset and tests on actual images. Extensive experiments on different infrared datasets demonstrate the effectiveness of the synthetic methods. The synthetically trained network obtains a mAP of 0.826, and the recall rate of small latent short circuit is superior to that of Faster R-CNN and U-Net, effectively avoiding short-circuit missed detection.
Highlights
In the metal electrorefining process, short circuits between electrodes cause the temperature of the electrodes to rise, the electrochemical reaction to stop, and the further reduction of the electrolytic efficiency (Aqueveque et al, 2009)
In order to overcome the difficulty of manually annotating a sufficient number of images and meeting the accuracy requirements of short circuit recognition tasks, we first propose a framework for automatically generating labeled images, and design an attention-based Faster R-Convolutional Neural Networks (CNN) for short circuit detection
This work focused on short circuit detection in infrared image of metal electrolysis scene with CNNs
Summary
In the metal electrorefining process, short circuits between electrodes cause the temperature of the electrodes to rise, the electrochemical reaction to stop, and the further reduction of the electrolytic efficiency (Aqueveque et al, 2009). In order to overcome the difficulty of manually annotating a sufficient number of images and meeting the accuracy requirements of short circuit recognition tasks, we first propose a framework for automatically generating labeled images, and design an attention-based Faster R-CNN for short circuit detection. To increase the detection accurancy, our detection scheme improves the Faster R-CNN by introducing an attention module This module fuses the semantic segment information of small-scale latent short circuits and the synthetic dataset, making the network focus on small objects during the extraction of features. 2) Improve the Faster R-CNN by introducing an attention module and design the short circuit recognition system for metal electrolysis, the system is trained only on synthetic samples and generalizes well to real images especially for the latent short circuit class.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have