改进的残差网络对红外图像热斑状态分类研究

Hot Spot State Classification of Infrared Image Based on Improved Residual

为提高光伏组件红外热斑图像识别准确率,提出一种基于多尺度残差和注意力机制相结合的新型卷积神经网络AMSRnet。在残差模块引入多卷积核,充分提取图像的深层特征信息,采用多层注意力模块,减少不必要的特征学习,增强特征的判别性,同时采用数据扩充方法防止模型过拟合。实验结果表明,AMSRnet模型训练识别自制的光伏组件红外图像热斑状态数据集,准确率高达95%,与VGG16等现有模型进行对比,AMSRnet模型的识别准确率比其他模型提高了4.41%~13.82%,且训练过程中准确率未出现明显的抖动现象,具有较高稳定性。

In order to improve the accuracy of infrared hot spot image recognition of photovoltaic modules,a new convolution neural network AMSRnet based on multi-scale residual and attention mechanism is proposed in this paper.The multiconvolution kernel is introduced into the residual module to fully extract the deep feature information of the image,and the multi-layer attention module is used to reduce unnecessary feature learning and enhance the discrimination of features.At the same time,the data expansion method is used to prevent the model from over-fitting.The experimental results show that the AMSRnet model trains to recognize the self-made infrared image hot spot state data set of photovoltaic modules,and the accuracy is as high as 95%.Compared with the existing models such as VGG16,the recognition accuracy of the AMSRnet model is 4.41% higher than that of other models,and there is no obvious jitter in the training process.