Research And Implementation Of Infrared Object Detection Based On Deep Learning

Infrared target detection is a basic task in the fields of infrared detection,intelligent security,and night driving assistance.Traditional infrared target detection algorithms rely on artificially designed features,and have the problems of low detection accuracy and poor migration capabilities.Deep learning has end-to-end feature expression capabilities and can extract robust semantic features.Therefore,research on infrared target detection methods based on deep learning is of great significance for improving target detection accuracy and model generalization capabilities.Target detection algorithms based on deep learning use shallow visual features such as color,texture,and geometry to build high-level semantic features.However,the fuzzy edges and poor contrast of infrared image will lead to the lack of shallow features,which will reduce the ability of feature representation,and lead to unsatisfied detection accuracy and detection failure in some scenes.Aiming at the above problems,this thesis starts with infrared image enhancement and target detection network optimization,combining with the engineering requirements of embedded infrared target detection,and carried out the following research work:In order to solve the problem of narrow dynamic range and the lack of details,an infrared enhancement algorithm based on filter layering is proposed.The algorithm achieves efficient image layering by reasonable approximation of bilateral filter,and achieves contrast enhancement and detail enhancement of infrared image by using improved MSR algorithm and adaptive masking enhancement algorithm.In view of the high computational complexity of yolov4,and the poor accuracy of bounding-boxes and the target-missed detection in infrared scene,an improved infrared target detection algorithm is proposed.The algorithm draws on the idea of cross-stage partial connection,designs a modular feature enhancement network to reduce the computational complexity,and improves the target detection accuracy by redesigning the border regression method and non maximum suppression strategy.Finally,in order to meet the engineering application requirements of infrared target detection,a real-time infrared target detection system based on NX embedded platform is designed.The system uses the QT image framework to build an easy-to-use graphic user interface,and uses the channel-prunning algorithm and TensorRT acceleration technology to reduce the model size,and doubles the inference speed when the accuracy is almost unchanged.Experiment results in this thesis show that the infrared enhancement algorithm can significantly improve the clarity and contrast of infrared image,and improve the detection accuracy;the improved yolov4 model can reduce the number of parameters by 17%and floating-point operations by 13%,and improve the detection accuracy of 1.5 percentage points;the embedded infrared target detection system can achieve the inference speed of 35.8fps,which can meet the requirements of real-time infrared target detection.