Substation Object Detection Based on Enhance RCNN Model

摘要

In the object detection task of substation, the low resolution object would suffer from serious information loss problem, so some low resolution objects with potential security risks cannot be detected by object detection models such as Faster RCNN. We combine Faster RCNN model with Wasserstein GAN model, and propose Enhance RCNN model especially for the low resolution object detection in the substation. In our model, discriminator in GAN is used to distinguish the abstract feature difference between the high resolution object and the low resolution object after supplementing feature. And generator is used to supplement the abstract feature for low resolution object, so that its feature distribution is consistent with the feature distribution of high resolution object, thus improving the overall detection effect. The experimental results show that for the typical object in the substation such as person, bicycle and vehicle, Enhance RCNN model averagely improves mAP (Mean Average Precision) and IoU (Intersection-over-Union) by 7.79% and 6.57% respectively when is compared with the other models including Faster RCNN, Fast RCNN and SSD. For the low resolution object whose ratio of the object pixel to total image pixel less than 0.2%, Enhance RCNN model averagely improves mAP by 10.44%.