多尺度融合超分辨率算法在无人机探测中的应用

Application of multi-scale fusion super-resolution algorithm in UAV detection

基于光电传感器的低慢小无人机探测系统能够快速准确地发现并识别无人机目标,但远距离非合作无人机目标在图像中像素比重过小,特征退化较明显,使识别率大大降低。图像超分辨技术能够从低分辨率目标图像区域中获得高分辨率图像并恢复更多的细节特征,现有超分辨技术很难在保证推理速度的前提下兼容图像的高低频特征,因此为了满足探测系统的需求,基于FSRCNN(fast super-resolution convolutional neural network)的特征提取与非线性映射网络结构并结合多尺度融合,提出一种包含4分支的轻量级多尺度融合超分辨率网络,能够在超分辨率图形中兼容高低频图像信息,且参数量较低,实时性高。经实验结果表明,该算法能够更加快速高效地重建出高分辨率的无人机轮廓与细节;在YOLOV3检测效果的实验中,该算法能够使无人机检测置信度平均提升6.72%,具备较高的实际应用价值。

The low-speed and small unmanned aerial vehicle(UAV)detection system based on photoelectric sensors can quickly and accurately find and identify the UAV targets.However,the proportion of pixels in the images of long-distance non-cooperative UAV targets is too small,and the degradation of characteristics is obvious,which greatly reduce the recognition rate.The image super-resolution technology can obtain the highresolution images from low-resolution target image regions and restore the more detailed features.The existing super-resolution technology is difficult to be compatible with the high and low frequency characteristics of images while ensuring the inference speed.In order to meet the requirements of detection system,based on the feature extraction and nonlinear mapping network structure of fast super-resolution convolutional neural network(FSRCNN),and combined with the multi-scale fusion,a lightweight multi-scale fusion superresolution network with 4 branches was proposed,which could be compatible with the high and low frequency image information in super-resolution graphics and with low parameter quantity and high real-time performance.The experimental results show that the UAV contours and details with high resolution can be reconstructed more quickly and efficiently by this algorithm.In the experiment of YOLOV3 detection effect,the confidence degree of the UAV detection can be increased by 6.72%by this algorithm,which has high practical application values.