基于clear-SSD的单点多盒飞机目标检测天气适用性

目前遥感影像目标检测算法大多针对良好天气，一旦出现雾霾，则必然影响检测效果。为使目标在良好天气或雾天条件下均能有较优的检测精度，提高模型适用性。以飞机检测为例，提出一种基于影像处理的clear-SSD单点多盒目标检测模型。该模型在SSD检测算法前增加了影像处理算法，即先对待检测的遥感影像进行清晰化处理，再通过SSD检测算法提取影像中的飞机。比较不同清晰化算法对检测精度的提升效果，选择适用性最优的算法作为模型前端，备选清晰化算法包括暗通道、高斯同态滤波及线性同态滤波，研究表明，三种清晰化算法对精度均有改善，其中高斯同态滤波的适用性最优，平均检测精度达到0.9843，比原始SSD模型提高了0.043，因此，将高斯同态滤波作为clear-SSD模型的影像处理部分。

Research on Weather Applicability of Single Shot MultiBox Aircraft Target Detection Based on clear-SSD

At present, most of the remote sensing image target detection algorithms are aimed at normal weather. Once haze appears, the detection effect will be inevitably affected. In order to make the targets are better detected under normal weather or foggy conditions, and improve the applicability of the model. Taking aircraft detection as an example, a clear-SSD Single Shot MultiBox Detector based on image processing was proposed. An image processing algorithm was added before the SSD detection algorithm, that is, the remote sensing image to be detected was cleared first, and then the aircraft in the image was extracted by the SSD detection algorithm. The improvement effect of different sharpening algorithms are compared on detection accuracy, and select the algorithm with the best applicability as the front end of the model. The alternative sharpening algorithms include dark channel, Gaussian homomorphic filtering, and linear homomorphic filtering. Research shows that the three sharpening algorithms all improve accuracy.The applicability of Gaussian homomorphic filtering is the best, and the average detection accuracy reaches 0.9843, which is 0.043 higher than the original SSD model. Therefore, Gaussian homomorphic filtering is used as the image processing part of the clear-SSD model.