Research On Infrared Image Object Detection Algorithm Based On Deep Learning

Infrared images are widely used in the military,firefighting,video surveillance and other fields,and object detection technology based on infrared images has significant research implications.Traditional infrared object detection algorithms will not only generate a large number of redundant areas,but also need to extract image features using artificially designed features.These disadvantages will make good detection results in complex infrared scenes challenging.With the development of deep learning,deep learning based infrared image object detection technology can improve detection accuracy and real-time performance.The main research problem in this paper is how to design a high-performance infrared image object detection algorithm and how to use semi-supervised learning to improve the detection accuracy of the algorithm.The work of this paper is divided into three parts:The first part is concerned with the design and optimization of a key point infrared image object detection algorithm.This paper proposes a CenterNet-based infrared image object detection algorithm.The algorithm is a keypoint-based single-stage object detection model.It has the advantages of not relying on the previous information of the anchor of dataset,flexible training,and no post-processing compared to other algorithms.Folowing that,we design a selfattention deconvolution network and a CIoU regression loss function to improve the detection accuracy and robustness of the algorithm.Finally,when tested on the FLIR infrared dataset,the mAP of the improved CenterNet infrared image object detection algorithm has increased by1.6% compared to the original algorithm,reaching 62.1%.The second part focuses on the development of a semi-supervised infrared image objcet detection algorithm based on key points.Semi-supervised learning uses a large amount of unlabeled data to improve the detection performance of the model and reduce the cost of making dataset.This paper proposes an IR image object detection algorithm based on CenterNet and Consistency based semi-supervised learning(IRCC).The detection accuracy of the semisupervised model is determined by the image enhancement technique used during the training process.In this paper,we design Object-based Image Mixture Enhancement(OMix),a novel image enhancement method,which is applied to semi-supervised object detection.Finally,using different scales of FLIR datasets,the IRCC model can effectively use unlabeled data to improve the detection ability of the algorithm.The IRCC model uses only 1% labeled data for training,and its mAP can reach 49.7%.The IRCC model based on OMix image enhancement improves its mAP by 1.5%.In the third part,an infrared image object detection system is designed and constructed for using in the intelligent fire helmet project.The system consists of two parts: the terminal and the server platform.The object detection algorithm is optimized using the MobileNet lightweight network and TensorRT to meet the real-time requirements of embedded terminals.The embedded terminal is controlled by the edge AI device,visualizes infrared images and sensor data in the near-eye display system,and support wireless communication via a 4G module.The server platform receives terminal information using the gRPC communication protocol,stores it in the MySQL database,and uses unlabeled infrared images to optimize the semi-supervised algorithm online.