基于深度学习的梯度聚类SSD算法参数选择

Parameter Selection of Gradient Clustering SSD Algorithm Based on Deep Learning

基于深度学习的航拍目标检测算法,由于硬件条件的限制,不能对大尺寸的航拍图像直接进行检测,通常采用滑窗法提取候选区域,但这种方式不能兼顾算法的实时性。一种梯度聚类SSD算法利用航拍图像中人造物体的边缘梯度进行区域建议,对建议区域利用SSD模型进行目标检测得到检测结果,能够一定程度地解决对大尺度航拍图像检测的实时性问题。本文根据航拍图像的特点,对基于梯度聚类SSD中梯度聚类算法进行参数选择,使其建议的区域更加准确。实验在DOTA数据集上对参数的效果进行了检验。

Due to the limitation of hardware conditions, the aerial object detection algorithm based on deep learning cannot directly detect large - scale aerial images. Usually, the sliding window method is used to extract candidate regions. However, this method cannot take into account the real - time of the algorithm. The gradient clustering SSD uses edge gradients of man - made objects in aerial images for clustering to obtain the region of interest of the image, which can solve the problem of large - scale aerial image detection in real - time to some extent. Then the the proposed areas are detected by SSD model to obtain test results. In this paper, according to the characteristics of aerial images, parameters of gradient clustering algorithm in the gradient clustering SSD are selected to make it more suitable for aerial image detection. The experiment tested the effect of the parameters on the DOTA data set.