Region-Free Cloud Recognition Based on Center Net

Abstract

Because of the significant influence produced by different types of cloud on photovoltaic power, accurate cloud identification is developed widely to predict the variation of photovoltaic power generation. Classical cloud-recognition methods generally use texture, structure, and color simultaneously to capture the cloud characteristic. In this paper, a Center Net-based cloud image recognition method is applied to cloud identification for the first time to simplify the operation and enhance the detection speed. The proposed algorithm adopts Hourglass Resdcn18, Resdcn101, and DLA -34 networks respectively to generate heatmap, and picks the heatmap peaks as a center point for target identification. Eight kinds of clouds are selected to verify the detection performance of the proposed algorithm: altocumulus, altostratus, cumulus, cirrus, cirrostratus, cumulonimbus, cumulonimbus, stratocumulus. The effectiveness of identification is compared under the structure of each network. The cloud center location and confidence are visualized under the DLA-34 model with higher confidence. The results showed that the highest confidence level of DLA-34 network is 0.973, which is higher than that of other networks apparently. Compared with Faster R-CNN, the cloud image recognition speed of Center Net is faster and the confidence is higher.