基于逐像素自适应对抗网络的电力巡检图像增强方法

Electric Power Patrol Inspection Image Enhancement Method with Per-Pixel Self-Paced Adversarial Network

针对电力智能巡检中图像细节丢失、边缘模糊,导致目标检测、识别错误的问题,提出一种基于逐像素自适应对抗网络(per-pixel self-paced adversarial network,PSPA)的超分辨率方法。该方法以生成对抗网络为基础,加入多重注意力机制,并通过逐像素比对的方式还原细节纹理。实验结果表明,该方法生成的超分图像不仅在人眼主观体验上优于现有算法,而且在电力设备巡检数据集测试中峰值信噪比和结构相似性指标对比其他最优结果分别提高了6.2和0.0993。对还原后的超分辨率图像使用Yolov3网络在无人机输电线路绝缘子数据集和电力施工安全帽佩戴数据集上进行目标检测,不仅能够降低漏检率,而且能够将检测置信度提高到接近原始高清图像的结果。

In view of the problem that the image details are lost and the edges are blurred in the intelligent patrol inspection of electric power,resulting in the wrong target detection and recognition,a super-resolution method based on per-pixel selt-paced adversarial network(PSPA)is proposed.This method is based on the generation of adversarial network,adds multiple attention mechanisms,and restores the detailed texture through pixel by pixel comparison.The experimental results show that the super-resolution images generated by this method are not only superior to the existing algorithms in human visual system,but also 6.2 and 0.0993 times higher than the existing algorithms in PSRN and SSIM.Then Yolov3 is applied on the super-resolution images recovered by different algorithms in the UAV transmission line insulator dataset and the power construction helmet wearing dataset.The experimental results demonstrate that the proposed method could not only decreases the residual error rate,but also improves the detection confidence as high as the high-resolution images。