基于深度学习的红外遥感目标超分辨率检测算法

Super-Resolution Infrared Remote-Sensing Target-Detection Algorithm Based on Deep Learning

红外遥感图像受限于红外衍射极限,其分辨率普遍较低,这为红外目标的精准检测和识别带来了困难。针对此问题,提出基于深度学习的红外目标超分辨率目标识别(SROD)算法,该算法主要包括两部分:第一部分是利用WDSR(Wide Activation for Efficient and Accurate Image Super-Resolution)对红外遥感图像进行超分辨率重建,将模拟传感器下采样方式处理的红外图像作为训练集。第二部分是基于Faster RCNN的目标检测,提出多尺度特征传递网络结构,将低层特征输入区域候选网络(RPN)层,降低了弱小目标像素被简化的概率,并利用可调节非极大值抑制方法,减少了对密集目标检测框的抑制作用。将该算法应用于整幅红外遥感图像,与相同训练集的Faster RCNN相比,目标检测的准确率提升了5.33%,召回率提升了12.22%,特别是小目标的召回率提升了13.25%。

Owing to the infrared diffraction limit,the resolution of infrared remote sensing images is generally low,which makes precise detection and recognition of infrared targets difficult.To address this problem,an infrared target super-resolution detection algorithm based on deep learning is proposed.The algorithm comprises two main parts.The first part implements Wide Activation for Efficient and accurate image super-resolution(WDSR)to reconstruct infrared remote sensing images,and uses infrared images processed by the downsampling method of the sensor as the training set.The second part involves target detection based on Faster region-based convolutional neural network(Faster RCNN).A multiscale feature transfer network structure is proposed.The low-level features are input to region proposal network(RPN),which reduces the simplification rate of weak and small target pixels.In addition,a nonmaximum suppression method is used to reduce the suppression of dense target detection frames.Compared with Faster RCNN using the same training set,the proposed algorithm increased target detection accuracy,the overall recall rate,and the recall rate of small targets by 5.33%,12.22%,and 13.25%,respectively.