首页|基于深度学习的实例分割边界框回归方法研究

基于深度学习的实例分割边界框回归方法研究

扫码查看
针对实例分割任务中图像中可能出现相互遮挡或边缘模糊导致边界框定位不准确的问题,本文提出了一种新的边界框回归损失函数。将边界框位置预测转化为估计定位置信度随位置变化的概率分布;考虑坐标点间存在联系,提出一种面积差计算方法;为了证明此方法可以很好地应用于先检测后分割的实例分割模型,本文使用Mask R-CNN作为基线。实验结果表明:在边界框检测及实例分割任务中,本文方法的精度优于其他方法,对于小物体的检测与分割效果更显著,训练和评估速度也更快。
A bounding box regression method of instance segmentation based on deep learning
A new function of bounding box regression loss in the instance segmentation task is proposed in this paper to solve the problem of inaccurate bounding box location caused by occlusion or blurring of edges in an image.Here,the location prediction of a bounding box is transformed into the probability distribution of the estimated posi-tioning confidence changing with position.A method for area difference calculation is proposed considering the rela-tion between the coordinate points.Then,to prove that the method can be well applied in the instance segmentation model of the detection followed by segmentation,mask R-CNN is taken as the baseline.Experimental results reveal that the proposed method outperforms other methods in terms of accuracy in bounding box detection and instance segmentation.Furthermore,the effect is more significant for small object detection and segmentation,and the train-ing and evaluation speed is faster than others.

computer visiondeep learningconvolutional neural networkinstance segmentationMask R-CNNbounding box regressionKullback-Leibler divergenceGaussian distribution

刘桂霞、吴彦博、李文辉、王天昊

展开 >

吉林大学 计算机科学与技术学院,吉林 长春 130012

哈尔滨工程大学 船舶工程学院,黑龙江 哈尔滨 150001

计算机视觉 深度学习 卷积神经网络 实例分割 Mask R-CNN 边界框回归 KL散度 高斯分布

国家自然科学基金项目国家自然科学基金项目

6177222661862056

2024

哈尔滨工程大学学报
哈尔滨工程大学

哈尔滨工程大学学报

CSTPCD北大核心
影响因子:0.655
ISSN:1006-7043
年,卷(期):2024.45(3)
  • 22