针对海洋牧场观测视频色彩失真和鱼类传统识别方法准确率低的问题,提出一种基于Faster-RCNN 的海洋牧场鱼类识别与分类方法.首先,由于海水环境的特殊性和复杂性导致观测视频图像品质差,采用SDI(Serial Digital Interface)信号色彩补偿系统来提高视频品质以此制作不同质量数据集;然后以Faster-RCNN为深度学习模型并提出优化特征提取网络与区域建议网络(RPN)来实现海洋牧场鱼类识别与分类.实验结果表明,该方法平均精度均值(Mean Average Precision,mAP)达到81.63％,与传统机器学习目标检测算法相比,显著提高了识别的准确率.
Identification andclassification of marine ranching fish based on the Faster-RCNN
Towards the problems of color distortion of the marine ranching observation video and low accuracy of traditional fish identification methods,a marine ranching fish identification and classifica-tion method based on Faster-RCNN was proposed.SDI(Serial Digital Interface)signal color compensa-tion system was used first to improve the poor video quality caused by the particularity and complexity of the marine environments,and then the optimized video was applied to produce data sets with diverse qualities.The Faster-RCNN works following a deep learning model with the feature extraction network and region proposal network(RPN)optimized and used to identify and classify marine ranching fish.The tentative experimental results showed that the mean average precision(mAP)of this method reached 81.63％,significantly improved the accuracy of recognition compared with traditional machine learning target detection algorithms.
fish identificationSDI signal color compensationdeep learningfaster-RCNN
NETL
NSTL
万方数据
