基于卷积神经网络的彩色图像声呐目标检测

Target detection of color image sonar based on convolutional neural network

随着人工智能和声呐成像技术的日益发展,利用卷积神经网络算法对声呐图像水下目标进行自动检测逐渐成为水声探测领域的重要研究方向之一。在公开数据集中,声呐的成像数据集比较稀少,于是采用本实验室开发的专利产品TKIS-I头盔式彩色图像声呐作为数据采集工具。为了保证检测的效果,采用了2018年在目标检测领域最佳效果的YOLOv3网络,并且通过目标预测框维度聚类、不同分辨率数据预训练网络、多尺度训练、多尺度预测与细微粒度特征结合等方法改进YOLOv3网络,使其能更好地进行声呐图像目标检测。除此之外,对声呐图像的预处理加入双线性插值算法,使得圆扫声呐对目标的表现更加清晰,易于检测,为以后水下机器人自动目标检测奠定基础。

With the development of artificial intelligence and sonar imaging technology, it has become one of the important research directions in the field of underwater acoustic detection to make the use of Convolutional Neural Network (CNN) algorithm to automatically detect underwater targets of sonar images. As the sonar imaging datasets are relatively rare in public datasets, the patented TKIS-I helmet-type color image sonar developed by our laboratory was used as data acquisition tool. In order to ensure the detection effect, YOLOv3 network with the best effect in the target detection field in 2018 was adopted and improved through the methods of dimension clustering of target prediction frame, data pre-training network of different resolutions, multi-scale training, multi-scale prediction and fine particle size feature combination to make it better to perform sonar image target detection. In addition, the bilinear interpolation algorithm was added to the preprocessing of the sonar image, which made the performance of target of the rounded sonar more clear and easy to detect, laying a foundation for the automatic target detection of underwater robots in the future.