基于改进Cascade-RCNN的内河航标检测算法

Inland Navigation Mark detection Algorithm based on Improved Cascade RCNN

摘要: 针对内河复杂环境下船舶智能航行对航标检测识别与分类技术的要求，通过对经典Faster-RCNN模型特征提取网络、锚框机制、检测框抑制算法、损失函数等进行优化，采用ResNeXt、Soft-NMS、GIoU等结构改进了级联Faster-RCNN模型——Cascade-RCNN网络。以长江中游武桥水道航标数据为例，开展了算法测试与验证。研究结果表明：基于改进Cascade-RCNN的目标检测算法综合性能最佳，平均精度均值约94.17%、用时190毫秒/帧。该算法能够有效适应内河航标目标较小、重叠、多样的特点，保持较高的精确度与召回率，可满足内河复杂通航场景下航标的检测精度与效率需求。

Abstract: The algorithm is developed to answer the request of intelligent navigation for technologies for detection, identification and classification of navigation marks. The feature extraction network, anchor box mechanism, detection window suppression algorithm and loss function of typical faster RCNN are improved and the structure of the cascade RCNN is modified by introduction of ResNext, soft-NMS and GIoU (generalized intersection over union). The algorithm is verified with the data of navigation Marks in Wuhan waterway. Experiments show that the algorithm achieved average accuracy of 94.17% with the process speed of 190 ms per frame. This algorithm is seen strong in dealing with overlapping small targets.