To improve the recognition accuracy of samples under uneven distribution, an improved ResNet50 convolutional neural network photovoltaic hot spot recognition algorithm is proposed. First, in order to increase the initial infrared texture information inflow and adjust the network width, a head grouping feature extraction module is designed and embedded into the residual network to improve the network' s extraction capability of image subtle features;Then, by combining channel attention mechanism with residual module, the weight of hot spot feature information between network channels is increased to improve model recognition performance and network con-vergence speed;Last, through data preprocessing methods such as image conversion into HSV color space and average H-component gradient histogram peak, negative samples are converted into multi classification datasets and used in the hot spot recognition network model to achieve visualization of the hot spot recognition results. Results show that compared to other algorithms,improved ResNet50 net-work significantly improves detection accuracy.
关键词
光伏热斑/图像识别/HSV颜色空间/ResNet50
Key words
photovoltaic hot spot/image recognition/HSV color space/ResNet50
维普
万方数据