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

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 high-resolution 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 super-resolution 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.