【作者】 赵鹏；

【导师】 陈海永；

【作者基本信息】 河北工业大学 ， 工程硕士（专业学位）， 2020， 硕士

【摘要】 光伏电池缺陷的存在不仅会影响其发电效率、产品质量,还会降低光伏电站发电的安全性。目前电池片的缺陷检测更多依赖于人工检测,存在效率低、工作强度大、成本高的缺点。因此,实现高效率、高精度的电池片裂纹检测具有重要的意义和价值。现如今,机器视觉的发展推动了智能缺陷检测技术的发展,受生产工艺、运输方式等因素的影响,光伏电池近红外图像的背景复杂,且裂纹缺陷表现出形状各异的多尺度特征,使得基于人工设计特征的传统机器视觉算法不再具有更强的通用性,检测精度也不够高。因此,本文从抑制复杂背景的干扰,降低误检提升准确率,提高模型对多尺度裂纹缺陷检测的适应性的角度,研究设计了三种基于深度学习的检测算法:为抑制近红外图像的非均匀纹理复杂背景对裂纹检测的干扰,提出基于密集连接卷积神经网络(Dense Net)和信息熵(entropy)数据融合策略的深度学习模型EntropyDense Net。其训练时的输入为128×128像素的图像块,模型与滑动窗口相结合实现了高于训练集像素大小的任何高分辨率图像的裂纹检测,充分利用了裂纹的空间结构,结合信息熵公式判断目标是否属于真性裂纹,完成以输入图像块为中心的结构预测。实验结果表明,所提模型有效地降低了裂纹缺陷的误检,提升了准确率。但是由于测试图像和裂纹具有不同的大小和比例,使得模型很难找到滑动窗口的最佳尺寸且Entropy-DenseNet评估图像的速度取决于裂纹的大小,性能有待进一步提升。为了直接获得高分辨率近红外图像的全局特征信息,提出了融合注意力CBAM的Faster-RCNN目标检测模型,实现对1024×1024像素图像的裂纹缺陷检测。模型采用VGG作为特征提取网络,将卷积注意力模块CBAM应用于RPN网络中,提升模型对裂纹缺陷的关注度,得到更加鲁棒的特征图。实验结果表明,融合CBAM的FasterRCNN性能高于原始检测模型,且优于仅采用通道注意力或空间注意力的FasterRCNN。但存在对复杂背景下的小尺寸裂纹的漏检。为提高模型对多尺度裂纹缺陷的适应能力,本文基于上述结果提出多尺度FasterRCNN目标检测模型。将更深的残差网络Res Net50与改进的路径聚合特征金字塔网络PA-FPN相结合,通过多层特征融合获取高分辨率、多尺度的包含丰富语义信息的特征图,提高模型在复杂背景下对小目标的特征表达能力。同时RPN网络采用损失函数Focal loss来降低训练过程中简单样本所占比重,使模型更关注难以区分的样本。此外,通过聚类算法k-means指导RPN设置更加接近实际裂纹尺寸的anchor,有利于目标框的位置回归。实验结果表明,所提模型获得了优异的多尺度裂纹检测效果。另外,本文基于光伏电站巡检的热红外图像数据集进行了实验,验证了多尺度Faster RCNN模型可以实现小目标缺陷高效且精准的检测,具有很好的鲁棒性与泛化能力。更多还原

【Abstract】 The existence of photovoltaic cell defects will not only affect its power generation efficiency and product quality,but also reduce the safety of photovoltaic power generation.At present,the defect detection of battery cells relies more on manual inspection,which has the disadvantages of low efficiency,high work intensity and high cost.Therefore,it is of great significance and value to realize high-efficiency and high-precision crack detection of photovoltaic cells.Nowadays,the development of machine vision promotes the development of intelligent defect detection technology.Due to factors such as production processes and transportation methods,the background of the near-infrared image of photovoltaic cells is complex,and crack defects show multi-scale features with different shapes.As a result,traditional machine vision algorithms based on artificial design features are no longer more generic,and the detection accuracy is not high enough.Therefore,in this thesis,from the perspective of suppressing interference from complex backgrounds,reducing false detection to improve accuracy,and improving the model’s adaptability to multi-scale crack defect detection,three research-based detection algorithms are designed:In order to suppress the interference of the non-uniform texture complex background of the near infrared image on crack detection,a deep learning model Entropy-Dense Net based on the densely connected convolutional neural network(Dense Net)and information entropy data fusion strategy is proposed.The input during training is an image patch of 128×128pixels.Combining Dense Net with a sliding window enables crack detection of any highresolution image larger than the pixel size of the training set.The model makes full use of the spatial structure of the crack and combines with the information entropy formula to determine whether the target belongs to a real crack,so as to complete the structural prediction centered on the input image patch.Experimental results show that the proposed model effectively reduces the false detection of crack defects and improves the accuracy.However,because the test image and the crack have different sizes and proportions,it is difficult for the model to find the optimal size of the sliding window and the speed of EntropyDense Net’s evaluation of the image depends on the size of the crack,so the performance needs to be further improved.In order to directly obtain the global feature information of high-resolution near-infrared images,a Faster-RCNN target detection model with attention CBAM is proposed to achieve crack defect detection for 1024×1024 pixel images.The model uses VGG as a feature extraction network,and the convolutional attention module CBAM is applied to the RPN network to increase the model’s attention to crack defects and obtain a more robust feature map.Experimental results show that the performance of CBAM-fused Faster-RCNN is higher than that of the original detection model,and it is better than Faster-RCNN that only uses channel attention or spatial attention.However,there is a missed detection of smallsized cracks in a complex background.In order to improve the adaptability of the model to multi-scale crack defects,this thesis proposes a multi-scale Faster-RCNN target detection model based on the above results.The model combines the deeper residual network Res Net50 with the improved path aggregation feature pyramid network PA-FPN.It is through multi-layer feature fusion to obtain highresolution,multi-scale feature maps containing rich semantic information,which can improve the model’s ability to express features of small targets in a complex background.At the same time,the RPN network uses the loss function Focal loss to reduce the proportion of simple samples in the training process,so that the model pays more attention to the samples that are difficult to distinguish.In addition,the clustering algorithm k-means guides the RPN to set the anchor closer to the actual crack size,which is beneficial to the position regression of the target box.Experimental results show that the proposed model achieves excellent multi-scale crack detection effect.In addition,this thesis has conducted experiments based on thermal infrared image data sets of photovoltaic power plant inspection.It is verified that the multi-scale Faster RCNN model can achieve efficient and accurate detection of small target defects,and has good robustness and generalization ability.更多还原