随着太阳能光伏发电技术的普及和应用范围的扩大，光伏发电系统的安全运行已成为电力系统发展的热点和焦点。光伏电池板作为电能转换的核心器件，由于其材料自身特性与外界环境干扰，导致其在工作中易存在断栅、裂痕、错位等缺陷，从而影响电池板的工作效率，甚至严重威胁光伏电站的运行安全性与稳定性。因此，设计一种准确、快速的光伏电池板缺陷检测算法至关重要。为此，本文开展基于深度学习与特征增强的光伏电池板缺陷图像检测算法研究。论文的主要研究内容包括：

（1）针对图像采集过程中由于外界复杂环境带来的干扰导致传统算法检测精度低的问题，提出一种基于亮度感知与双通道分层的光伏EL图像缺陷检测算法。首先，为改善图像在光线较暗情况下存在边缘轮廓模糊的不足，设计一种基于亮度感知的图像增强模块，以增强原始图像中异常缺陷部分的显著度；其次，通过金字塔变换对增强后的图像进行多尺度分解得到高、低频图像，并利用“最大绝对值”规则和计算权重系数进行融合，进一步提高图像质量；最后，在YOLOv8主干网络中通过结合通道与空间注意力中所提取的目标特征，增强检测算法获取缺陷位置与特征信息的能力。实验结果表明，与9种经典检测算法相比，所提算法具有明显优势，其中mAP@.5、准确率和召回率分别达到了88.3%、88.6%和86.3%。

（2）针对在光伏电池板异常缺陷目标检测时存在难以兼顾检测精度与训练复杂度的问题，提出一种基于接收场块与全局协调的光伏EL图像缺陷检测算法。首先，为减少网络出现漏检与误检的现象，通过在YOLOv8基础网络中采用精确快速目标检测的接收场模块，增强异常缺陷目标区域的特征表达能力；其次，构建并引入全局协调注意力映射模块，提高对目标的识别与检测能力；最后，将YOLOv8中的Conv卷积模块替换为GhostConv模块，以减少网络模型的参数量。通过与13种经典检测算法进行对比，结果表明，所提出算法能够有效改善检测算法的性能，其中每张图片的检测速度可达到14.1ms，同时准确率、召回率和mAP@.5三项指标可分别达到85.4%、91.4%与93.1%。

With the popularization of solar photovoltaic power generation technology and the expansion of application scope, the safe operation of photovoltaic power generation system has become the hot spot and focus of power system development. Photovoltaic panels as the core device of electric energy conversion, due to its material characteristics and external environment interference, it is easy to have defects such as broken grid, cracks, dislocation in the work, which affects the efficiency of the panels, and even seriously threatens the operation safety and stability of photovoltaic power stations. Therefore, it is very important to design an accurate and fast defect detection algorithm for photovoltaic panels. For this reason, this thesis studies the defect image detection algorithm of photovoltaic panels based on deep learning and feature enhancement. The main research contents of this thesis include:

(1) In order to solve the problem of low detection accuracy of traditional algorithm due to interference caused by complex external environment in the process of image acquisition, a defect detection algorithm of photovoltaic EL image based on brightness perception and double-channel layering is proposed. Firstly, an image enhancement module based on brightness perception is designed to enhance the saliency of abnormal defects in the original image, in order to improve the lack of blurred edges in the image under low light. Secondly, high and low frequency images are obtained by multi-scale decomposition of the enhanced images through pyramid transformation, and the "maximum absolute value" rule and calculated weight coefficient are fused to further improve the image quality. Finally, the ability of the detection algorithm to obtain defect location and feature information is enhanced by combining the target features extracted from the channel and spatial attention in the YOLOv8 backbone network. Experimental results show that the proposed algorithm has obvious advantages over 9 classical detection algorithms, and mAP@.5, accuracy rate and recall rate reach 88.3%, 88.6% and 86.3%, respectively.

(2) Aiming at the difficulty of both detection accuracy and training complexity, a defect detection algorithm based on receiving field block and global coordination for photovoltaic EL image is proposed. First, in order to reduce the phenomenon of missing detection and false detection in the network, the receiving field module of accurate and fast target detection is adopted in the basic network of YOLOv8 to enhance the feature expression ability of the abnormal defect target region. Secondly, a global coordination attention mapping module is constructed and introduced to improve the ability of target recognition and detection. Finally, the Conv convolution module in YOLOv8 is replaced with the GhostConv module to reduce the number of parameters in the network model. Compared with 13 classical detection algorithms, the results show that the proposed algorithm can effectively improve the performance of the detection algorithm, in which the detection speed of each image can reach 14.1ms, and the accuracy rate, recall rate and mAP@.5 can reach 85.4%, 91.4% and 93.1%, respectively.