在光伏产业中,光伏板缺陷的检测至关重要,针对目标检测算法SSD在光伏板缺陷检测中精度不高、速度不快、存在漏检、误检等问题,本文提出了一种改进的SSD算法,主要从三个方面改进:首先,使用融合了ECA注意力机制的Resnet50网络替换原SSD中的VGG-16网络,减少了计算量,提高了检测精度;其次,使用Involution算子替换原SSD中的Conv7卷积层,使模型更加轻量;最后,引入Focal Loss损失函数解决正负样本不平衡的问题.经实验分析,改进的SSD算法相较于原SSD算法的mAP提升了4.41个百分点,同时检测速度提升了6.55frame/s.
In the photovoltaic industry,it is very important to detect the defects of photovoltaic panels.In view of the problems of low accuracy,low detection speed,missed detection and false detection in the target detection algorithm SSD,this paper proposes an improved SSD algorithm,which is mainly improved from three aspects:Firstly,the VGG-16 network in the original SSD is replaced by the Resnet50 network with ECA attention mechanism,which reduces the calculation amount and improves the detection accuracy;Secondly,the Conv7 Involution layer in the original SSD is replaced by the involvement operator,which makes the model lighter.Finally,the Focal Loss function is introduced to solve the problem of unbalanced positive and negative samples.The experimental analysis shows that the improved SSD algorithm improves the mAP by 4.41 percentage points compared with the original SSD algorithm,and the detection speed is improved by 6.55 frame/s.