基于孪生网络和长短时记忆网络的输油管道泄漏检测方法

Oil leakage detection based on siamese network and LSTM

针对目前大部分基于图像的输油管道泄漏检测方法都是基于变化检测的,并很容易受到噪声、光线变化等因素的影响产生误报的问题,提出了基于孪生网络和长短时记忆网络的输油管道泄漏检测方法。首先对视频中的每一帧图像,使用孪生网络进行变化检测以检测出可能的油滴区域。由于变化检测受阈值的影响,容易出现漏检或误检,因此降低变化检测的阈值,尽可能地检测出所有可能的变化区域。然后采用长短时记忆(LSTM)网络建模每个变化区域的变化趋势来进行泄漏检测,有效地降低误检。在王家沟油库的测试结果表明,与背景减除法+阈值和孪生网络+阈值相比,所提出的方法可以有效降低误检率(由20%降为1%)、提高正检率(由80%提高为99%)。该方法可以应用在输油管道的自动检测中,有效地对输油管道的状态进行监控。

Most current image based oil leakage detection methods are based on change detection, and are easily affected by noise, illumination change and other factors, which cause false positives. To solve the problem, an oil leakage detection method based on siamese network and LSTM (Long Short-Term Memory) was proposed. Firstly, siamese network was employed to detect change regions possibly related to oil leakage in every frame in surveillance video. As change detection was affected by the threshold, a low threshold was used to detect the possible change regions as many as possible. Then, LSTM was used to model the change tendency of each change region to predict whether the region was related to oil leakage. The experiment results on Wangjiagou oil depot indicate that the proposed method can detect oil leakage with higher precision (99% vs. 80%) and lower false positive rate (1% vs. 20%), compared to the background subtraction method with threshold and the siamese network with threshold. This method can be applied to automatic detection of oil pipelines effectively.