Smart UAV Infrared Image Defect Detection using YOLOv3 and Gaussian Mixture Models

Abstract

To improve detection performance, a model with strong universality and adaptability to different environments was established. This study proposed a migration and optimization method for drone infrared image defect detection and recognition models based on deep learning and transfer learning. This method proposed a temperature prediction method for infrared images based on fully convolutional neural networks to reduce storage and transmission costs. In addition, the study introduced a pixel level infrared image defect segmentation method based on Gaussian mixture model to solve the background interference problem caused by large defect area angles. Experiments confirmed that the algorithm performed well in depth detection, achieving an accuracy of 90% and processing time of only 11 ms. Compared to other methods, this algorithm not only had a low false detection rate of 2.6%, but also reduced the running speed by 63 ms and 27 ms compared to the wavelet texture feature statistical analysis and the spectral residual visual saliency methods. In addition, through filtering and morphological processing, this algorithm provided a more complete defect area, almost eliminating noise, which was helpful for further defect area analysis. This has important practical significance for further improving the application of drone infrared image defect detection.

Keywords

• detection
• gaussian mixture model
• infrared images
• transfer learning
• yoloV3