基于图像分割和密度聚类的遥感动目标分块提取
Image segmentation and density clustering for moving object patches extraction in remote sensing image

摘要 大幅宽遥感图像的动目标检测研究中，卷积神经网络虽然取得了显著效果，但算法存在目标搜索空间庞大、模型极其消耗时间及计算资源的问题，因此本文从目标区域预筛选的角度给出了针对性优化方法。首先，基于局部误差处理的策略，改进了现有的图像分割算法来粗糙地提取动目标可能存在的区域。然后，以相邻区域合并、减少总数量和面积为目的，设计了一种基于空间约束的密度聚类算法——SC-DBSCAN，其以分治思想来降低问题的规模，通过空间尺寸的先验约束自适应地将数据划分为多个相互独立的簇，并针对簇的复杂程度选择相应的合并策略，在复杂簇中，考虑到合并结果与对象遍历顺序相关，易陷入局部最优，引入基于模拟退火思想的随机扰动有效提升了输出的图像块质量。最终，通过减少模型推断次数及避免目标的重复检测，显著地改进动目标检测的整体效率。
Abstract：Recently, moving object detection in large-scale remote sensing images achieves outstanding performance by fully convolutional neural network. However, handling such data is very time-consuming because the search space is extremely large. This paper proposes a specific improved method from the point of candidate region proposals. First, irregular candidate areas are roughly extracted by neighborhood differencing and local errors handling. Then a spatial-constraint based density cluster algorithm (SC-DBSCAN) is proposed to merge adjacent areas into patches as CNN input, which aims to reduce final outputs' amount and area. Through the priori of space constraints, this algorithm can adaptively divide data into multi types of clusters, and choose different merging strategies according to the complexity of clusters. For complicated clusters, the outputs are related to traverse sequence of each object, and thus a random search strategy based on simulated annealing is applied to avoid local optima and improve the patches' quality. Finally, by reducing the times of model inferences and avoiding redundant object detections, the detection efficiency of proposed method is significantly improved.