一种低功耗的电网巡检故障图像识别方法

A fault image recognition method with low power consumption for power grid patrol inspection

针对计算机平台在图像处理过程中面临的高能耗问题,研究了电网巡检故障图像识别的低功耗神经网络加速方法.采用直接映射方式将卷积层与全连接层的计算分别映射至独立的计算核心,提出了针对图像处理过程不同阶段的优化方案,实现了不同运算层次与硬件资源之间的匹配,并通过遗传算法得到了神经网络的并行优化参数.结果表明,优化后的LetNet-5与AlexNet卷积神经网络运行能效分别为优化前的1.94倍和1.8倍,单张图片的平均处理速度约为原来的4~5倍和2~3倍.

Aiming at the problem of high energy consumption in the image processing of computer platform,a neural network acceleration method with low power consumption for fault image recognition during the power grid patrol inspection was studied.The computation of convolution layer and full connection layer were mapped to independent computing cores by a direct mapping mode,respectively.The optimization schemes for different stages of image processing were proposed,and the matching among different computation levels and hardware resources was realized.In addition,parallel optimization parameters of neural network were obtained by a genetic algorithm.The results show that the energy efficiencies of optimized LetNet-5 and AlexNet convolutional neural networks are 1.94 and 1.8 times that before optimization,respectively.Furthermore,the average processing speeds of a single picture are about 4 to 5 times and 2 to 3 times higher than that before optimization,respectively.