Research On Robust Cloud Classification Of Ground-Based Image Based On Deep Learning

With the growth awareness of environmental protection and the demand for clean energy,a clean solar energy has attracted wide attention.However,the output of solar energy is greatly affected by the weather condition.The main factors are the type and position of cloud.To address these issues,we focus on the classification and segmentation of ground-based cloud images using deep learning and compressed sensing techniques:(1)A two channel convolution neural network classification algorithm is proposed.We use transfer learning to train two sub-networks features.The fused feature experimental results show the accuracy improvement of about 1.3% compared to the traditional algorithm.(2)A robust classification method is proposed for the problem of interference and occlusion.The features of the cloud are extracted from multiple sub-networks and fused according to the variances.The ground-based cloud image is classified by the weighted sparse representation algorithm.The experimental results show that the accuracy of the proposed method is 93.28% with 25% occlusion,while the traditional method is75.15%.(3)A ground-based cloud segmentation method based on deep learning is proposed for image segmentation.A sub dataset is constructed from existing ground-based cloud dataset.The experimental results show that Deeplabv3+ model can effectively improve the stability and accuracy of segmentation.(4)According to the demand of the project,the algorithm is deployed on the RK3568 after environmental configuration and model conversion.User interface is designed for human-machine interaction.