基于RepGT损失的改进Faster R-CNN的包裹检测算法

Improved Faster R-CNN Parcel Detection Algorithm Based on RepGT Loss

针对目前我国快递包裹中转中心面临的快递包裹文件数量密集型问题以及包裹辨识检测算法技术,提供了一个经过改进的Faster R-CNN辨识检测算法。通过修改Faster R-CNN算法中的损失函数,用RepGT损失函数替代原回归项中的损失函数,使得选取的包裹候选框更接近包裹目标框,完成图像检测。通过数据实验发现,改进后的算法比传统的Faster R-CNN在精度上提升了2.38AP,同时发现当损失函数中参数σ=1时,检测精确度达到最高。

In this paper,aiming at the problems faced by the express parcel transfer centers in China,we proposed an improved Faster R-CNN identification and detection algorithm.Next,by modifying the loss function in the Faster R-CNN algorithm and replacing the loss function in the original regression with the RepGT loss function,we arrived at parcel identification frames closer to the target frames to complete the image detection.At the end,through a numerical example,it is found that the improved algorithm is 2.38AP more accurate than the traditional Faster R-CNN algorithm,and also identified the parameterization of the loss function to realize the highest detection accuracy.