融合视觉特征的光伏组件语义分割模型研究

SEMANTIC SEGMENTATION MODEL OF PHOTOVOLTAIC MODULES BASED ON VISUAL FEATURES

针对光伏组件红外图像的分割问题,使用MobileNetv2作为DeepLabv3+的主干特征提取网络并使用位置通道注意力模块减少背景干扰,引入混合条带池化对ASPP模块进行优化,帮助模型进一步捕获全局和上下文信息。针对检测困难的屋顶光伏组件设计DeepLabv3-T网络,在上述改进的基础上融入纹理信息进行选择性背景抑制,实现光伏组件的精确分割。在PV_large和PV_roof数据集上进行实验证明该文方法优于现有技术,DeepLabv3-T相较于DeepLabv3+,mIoU值分别提高了2.74%和10.53%。此外,设计消融实验表明各个改进模块的有效性。

To solve the segmentation problem of infrared images of photovoltaic modules,MobileNetv2 is used as the backbone feature extraction network of DeepLabv3+and the location channel attention module is used to reduce background interference.Mixed strip pooling is introduced to optimize the ASPP module,which helps the model to further capture global and contextual information.The DeepLabv3-T network is designed for rooftop PV modules with difficult detection.Based on the above improvements,texture information is incorporated into the selective background suppression to achieve accurate segmentation of PV modules.Experimental results on the PV_large and PV_roof datasets demonstrate that the text-based approach is superior to the prior art,and the mIoU value of deeplabv3-t is 2.74%and 7.93%higher than that of DeepLabv3+,respectively.In addition,ablation experiments are designed to demonstrate the effectiveness of each improved module.