Object Detection Based on Feature Scale Fusion and Feature Scale Enhancement

Abstract

Listen

Translate article

Currently, object detection is widely used to deal with the image analysis problem, which is the most important task in computer version. As the high image resolution leads to high computational cost and the low image resolution leads to low accuracy, the key bottleneck of CNN based object detection is the image resolution selection. In this work, we solve the problem of obtaining powerful object detection effect with feature scale fusion and feature scale enhancement. The proposed module can achieve significant accuracy by applying Feature Scale Fusion Module (FSFM) and Feature Scale Enhancement Module (FSEM) to the feature extraction layer of Faster R-CNN. In this case, the enhanced feature map become the input of RPN layer and ROI layer of Faster R-CNN. The accuracy gain of the propose module is verified via the Pascal VOC and MSCOCO datasets, which is proved to obtain significant improvements over the most advanced detection models.