基于改进U-Net网络的光伏板图像分割方法

Photovoltaic panel image segmentation method based on improved U-Net network

光伏板区域识别与分割对光伏板的缺陷精确检测和组件精准定位有重要意义.在复杂环境下,针对光伏板图像存在对比度不强、边界模糊、背景复杂等影响分割的问题,提出了一种改进U-Net网络的光伏板图像分割方法.首先,搭建基于U-Net网络的对称编码-解码结构骨干网络;其次,使用深度可分离卷积替代传统卷积,并将高效ECA注意力模块添加到两组深度可分离卷积之间,以两组深度可分离卷积和一个ECA注意力模块组成一个block块,利用多个block块提升多层网络的分割性能;之后,引入交叉熵损失、Dice损失、Focal损失线性加权和作为新的损失函数,训练改进U-Net网络;最后,为验证方法的有效性,将改进U-Net网络与MobileNetV2网络、U-Net网络、Res-U-Net网络分别在3 200张光伏板红外图像数据集上进行横向对比.结果表明:改进U-Net网络的PA值和MIoU值达到了0.993 1和0.980 2,均优于其他3种网络模型,且参数量只有U-Net网络和Res-U-Net网络的33.3%和30.4%,仅次于MobileNetV2网络.因此,改进U-Net网络具有较高的准确性和泛化性,能够完成光伏板图像分割任务.

Photovoltaic panel region recognition and segmentation is of great significance to defect detection and module positioning.Aiming at the problems that affect the segmentation of photovoltaic panel image,such as weak contrast,fuzzy boundary and complex background,it proposes a photovoltaic panel image segmentation method based on improved U-Net network.Firstly,the backbone network of symmetric encoding-decoding structure based on U-Net network is built.Secondly,depthwise separable convolutions are used to replace traditional convolutions,and an efficient ECA attention module is added between the two sets of depthwise separable convolutions.Two sets of depthwise separable convolutions and an ECA attention module form a block,and multiple blocks are used to improve the segmentation performance of the multi-layer network.Then,the linear weighted sum of cross entropy loss,Dice loss and Focal loss is introduced as new loss functions to train and improve U-Net network.Finally,to verify the effectiveness of the method,the improved U-Net network was compared with MobileNetV2 network,U-Net network and Res-U-Net network on 3200 photovoltaic panel infrared image datasets respectively.The results show that the PA value and MIoU value of the improved U-Net network reach 0.9931 and 0.9802,which are better than the other three network models.The number of parameters is only 33.3%and 30.4%of U-Net network and Res-U-Net network,second only to MobileNetV2 network.Therefore,the improved U-Net network has high accuracy and generalization,and can complete the task of photovoltaic panel image segmentation.