面向输电线路异常目标识别的模型加速技术

Model Acceleration Technology for Abnormal Target Recognition of Transmission Lines

中文关键词:CNN加速  目标识别  YOLOv3  RISC-V

keywords:CNN acceleration  target recognition  YOLOv3  RISC-V

中文摘要:针对输电线路目标识别帧率较低的问题，提出一种改进的YOLO（you only look once，YOLO）v3-tiny网络模型。通过设计基于深度卷积可分离的新型网络架构，并引入正负样本置信度以优化损失函数，显著提升了所提模型的检测速度；采用像素自适应和通道自适应技术进行图像特征融合，对激活函数通过寄存器移位操作来加速运行，实现了网络模型参数与检测精度之间的平衡。提出一种硬件移位逼近浮点数的移位方法，简化了硬件运算，并基于平均准确率提出硬件移位准则，最终确定了适用于存内计算的卷积神经网络加速器方案。该方法在降低功耗38.9%和减少逻辑资源消耗83.6%的同时，平均准确率仅下降0.1%；在电力数据集上，实现了70.4%的平均识别准确率。该文的创新之处在于对网络架构和硬件加速方案进行优化，显著提高了目标识别的效率。

Abstract:This paper proposes an improved YOLOv3-Tiny network model to address the low frame rate in transmission line target recognition. By designing a novel network architecture based on depth wise separable convolution and optimizing the loss function with positive/negative sample confidence， the detection speed of the proposed method is substantially increased. Furthermore， the use of pixel and channel adaptive techniques for image feature fusion， along with the acceleration of the activation function through register shift operations， achieves a balance between network model parameters and detection accuracy. The paper also presents a hardware shift approximation method for floating-point numbers， which simplifies hardware operations， and introduces hardware shift criteria based on average accuracy. Ultimately， a convolutional neural network accelerator scheme tailored for in-memory computing is established. This scheme reduces power consumption by 38.9% and logic resource consumption by 83.6%， with only a slight decrease in average accuracy of 0.1%. On the power dataset， this approach has realized an average recognition accuracy of 70.4%. The innovation of this paper lies in the optimization of the network architecture and hardware acceleration scheme， which markedly enhances the efficiency of target recognition.