【摘要】 视频监控作为海漂垃圾定点监测的一种重要手段,具有可长时间、连续获取某一重点区域海漂垃圾动态变化过程的特点。为了实现海漂垃圾的自动化识别,从而连续监测厦门湾海漂垃圾的变化情况,本文以厦门嵩屿码头视频监控获取的海漂垃圾原始图片为样本,构建了基于VGG16的卷积神经网络判别模型。设计了4种训练方案实现橘黄色条带状物质(可能是由船舶排污产生的泡沫与悬浮泥沙混合物形成的)和木屑类垃圾的分类识别。结果表明当提取目标的像素点占总像素点的比例越高,模型越可以得到充分的训练,从而得到较高的分类精度。当海漂垃圾占比只有2%～3%时,该方法的分类提取精度仍可达90%以上。同时,该方法能有效减少干扰物像素点与垃圾像素点的混淆,具有很强的实用性。更多还原

【Abstract】 Video surveillance, an important means of monitoring floating marine debris, has the characteristics of long-term and continuous acquisition of the dynamic change process of such debris in a key area. In this paper, using original images acquired by video surveillance of Songyu Wharf, Xiamen Bay, China, a convolutional-neural-network(CNN) model based on VGG16 was constructed to identify floating marine debris. Four training schemes were designed to discriminate two classes of such debris,including orange strips(may be formed by a mixture of suspended sediment and foam produced by ship discharge) and wood chips.Results show that the higher the proportion of target pixels to be classified, the more fully the CNN model will be trained and the higher the classification accuracy. Even if the proportion of floating marine debris was only 2%–3%, the classification accuracy reached up to 90%. Furthermore, this model can effectively reduce the confusion between interference pixels and targeted pixels,which endows it with powerful practicability.更多还原