基于元学习的少样本水闸图像识别方法研究

Research of few-sample sluice image recognition method based on meta-learning

针对实际工程环境中采集的水闸图像样本不均衡、前后景混融导致的识别效果不好的问题,提出一种基于元学习的少样本水闸图像识别方法。首先构建水闸图像数据集,并使用图像增强和预处理对数据集进行优化;再使用多头注意力,提升网络准确捕捉多种与任务相关的关键特征信息的能力,更好地与时序卷积协作,进一步提高水闸图像的识别效果。在构建的sluice-ImageNet数据集上进行实验,实验结果表明,相比其他方法,所提方法在水闸启闭状态图像识别任务上更具有效性和优越性。该方法部署于重点水利工程视频监测平台,辅助人工监管,可实现对水闸异常运行情况的实时监测,为防汛决策提供智能化支持。

Aiming at the problem of poor recognition effect caused by uneven samples of sluice gate images collected in the actual engineering environment and the mixing of front and rear views,a few-sample sluice image recognition method based on meta-learning was proposed.First,the sluice image dataset was constructed,and the dataset was optimized using image enhancement and preprocessing.Then,multi-head attention was used to improve the ability of the network to accurately capture a variety of key feature information related to the task,and better cooperate with the temporal convolution to further improve the recognition effect of the sluice image.Experiments were carried out on the constructed sluice-ImageNet dataset.The experimental results show that the proposed method is more effective and superior than other methods in the sluice opening and closing recognition task.The method was deployed on the video monitoring platform of key water conservancy projects to assist manual supervision and realize real-time monitoring of abnormal operation of sluices,and provide intelligent support for flood control decision-making.