| 基于改进Vision Transformer的光伏电池缺陷识别研究 |
投稿时间：2022-08-08 修订日期：2022-12-07  点此下载全文 |
| 引用本文： |
| 摘要点击次数: 76 |
| 全文下载次数: 0 |
|
|
| 中文摘要:光伏电池片是太阳能发电系统的核心组件。电池片的磨损、裂纹等缺陷不仅影响电池寿命，同
时降低了电能转化效率，而传统的人工缺陷检测方法耗时长效率低。本文设计了一个基于 CNN 和视觉
Transformer（ViT）的光伏电池缺陷检测模型。首先，针对 ViT 对输入图像进行分割造成无法感知图像全
局信息的问题，设计了一个包含 12 层卷积的残差网络，并与特征金字塔结合，获取输入图像不同尺度的特
征，并将特征图分割池化后作为 Transformer 的输入信息。其次，针对 Transformer 手动设计位置编码函
数的不足，设计了一个位置编码分支模块，用以实现位置自编码。在电池片缺陷图像数据上的实现结果表
明，提出的模型在不增加计算量的情况下，提升了准确度，证明了模型的有效性 |
| 中文关键词:Defect Recognition of photovoltaic cells based on Convolutional Neural Network and Scaled vision Transformer |
| |
| Defect Recognition of photovoltaic cells based on Convolutional Neural Network and Scaled vision Transformer |
|
|
| Abstract:Photovoltaic cells are the core component of solar power generation system. The
defects such as wear and crack of the cells not only affect the battery life, but also reduce the
energy conversion efficiency. Traditional manual defect detection method is time-consuming and
inefficient. This paper designs a defect detection model of photovoltaic cells based on CNN and
scaled Vision Transformer (ViT). Firstly, since ViT cannot perceive the global information of the
image due to the segmentation of the input image, a residual network containing 12 convolution
layers is designed, which is combined with the feature pyramid network to obtain the features of
different scales of the input image, and the feature map is segmented and pooled as the input
information of Transformer encorder. Secondly, since the deficiency of Transformer position
coding function designed manually, a position coding branch module is designed to realize
position self-coding. The Experiment results on the defect image dataset show that the proposed
model improves the accuracy without increasing the amount of calculation. |
| keywords:Convolutional Neural Network,Attention Mechanism,Transformer, Defective Detection |
| 查看全文 查看/发表评论 下载pdf阅读器 |
次