|
The detection accuracy is low in the existing deep learning detection to extract solar PV panels sticky, incomplete and under the interference of factors such as roofs and substations in four different scenarios: water, saline, low-density grass and fields. In this study, U-Net neural network is improved by adding attention mechanism and residual network. A deep learning network model for solar PV panels detection that incorporates attention mechanism and residual structure: UNet Pro, is established. This model, based on U-Net, is able to pay more attention to the detection of solar PV panels by incorporating up-sampling and down-sampling in the process of attention mechanism. It is able to pay more attention to the effective detailed information and exclude the interference factors. Meanwhile, the residual module is embedded after the Relu function in each layer of the U-Net neural network. It effectively avoids the problems of gradient explosion, feature loss and network degradation. For solar PV datasets in different scenarios, the effects of adding the attention mechanism and the residual structure have been improved to different degrees. Compared with U-Net, mF1 value can be improved up to 94.67% and at most 4.46%. mIoU can be improved up to 93.75% and at most 4.74%.
|
