基于多分辨率特征自选择的遮挡物识别算法

光伏组件的遮挡物识别是光伏运维系统中不可或缺的环节，传统识别算法多依赖人工巡检，成本高昂且效率低下。基于卷积神经网络，提出了一种面向光伏组件的遮挡物识别算法PORNet。通过引入特征金字塔，构建多个分辨率下具有丰富语义信息的图像特征，提升对遮挡物尺度和密度的敏感性。通过特征自选择，筛选出语义最具代表性的特征图，以加强物体环境的语义信息表达。用筛选出的特征图完成遮挡物识别，从而提升识别准确率。在自建光伏组件落叶遮挡数据集上进行了实验比较和分析，并对识别性能进行了评估，通过与现有物体识别算法相比，所提算法的准确率和召回率分别提升了9.21%和15.79%。 

Occlusion recognition algorithm based on multi-resolution feature auto-selection

The identification of obstructions of photovoltaic modules is an indispensable link in modern photovoltaic operation and maintenance systems. Traditional identification methods mostly rely on manual inspections, but they are costly and inefficient. Therefore, based on the convolutional neural network, PORNet, an occlusion recognition algorithm for photovoltaic modules, is proposed. By introducing feature pyramids, image features with rich semantic information at multiple resolutions are constructed, enhancing the sensitivity to the scale and density of occlusions. Through feature auto-selection, the most representative feature maps are screened out to strengthen the semantic information expression of the object contexts. Finally, the screened feature map is used to complete the occlusion recognition, improving the recognition accuracy. Experimental comparison and analysis are carried out on the self-built photovoltaic module falling leaf occlusion dataset, and the recognition performance is evaluated. Compared with existing object recognition methods, the accuracy and recall rate of the proposed method are increased by 9.21% and 15.79%, respectively.