基于无监督学习卷积神经网络的声呐图像分割

Sonar Images Segmentation Based on Unsupervised Convolutional Neural Network

摘要:

声呐图像分割是图像分割技术发展中的组成部分，是水下目标识别与检测的重要一环。传统方法中基于有监督分割方法的算法往往代价较大，表现出试验周期长，实时性较差，运行速率较慢等不足。并且由于声呐图像的成像质量差、分辨率不高、边缘条件不清晰、人工标注工作量大等客观因素，不易建立用于有监督模型训练的大规模数据集，使得传统分割方法越来越不适应当前实际应用的多方面要求。本文将基于无监督学习卷积神经网络引入到声呐图像分割任务中，分割模型通过对单帧声呐图像进行训练和测试，最后经过推理得到将阴影区和目标高亮区分割后的声呐图像，得到分割出来的水下目标。通过对实验的分割结果进行各项指标分析，证明此方法有着更好的运行效率和分割精度，并且实时性较高，综合性能优于传统方法。

Abstract:

Sonar image segmentation plays a very significant role in the development of image segmentation technology. Due to the factors such as poor imaging quality, low resolution and unclear edge conditions of sonar images, it is difficult to screen and obtain large-scale target data sets, and the cost of manual annotation is high from time to time. The segmentation algorithm based on supervised segmentation method is often doing with high cost, including long test period, poor real-time performance, and high operation cost. In this paper, the unsupervised learning convolution neural network (UCNN) is introduced into the sonar image segmentation task. The segmentation model trains and tests by using one single frame sonar image, and finally obtains the sonar image after segmenting the shadow area and target highlight area through reasoning process. Through the analysis of the experimental results, it is proved that this method has good operation efficiency and accuracy, and has high real-time performance too.