一种结合多尺度特征融合与像素损失加权的显著性目标检测方法

A saliency target detection method combining multi-scale feature fusion and pixel loss weighting

在显著性目标检测领域,常使用二分类交叉熵函数计算损失,但该损失函数独立计算各像素点的信息,忽视了每个像素之间的联系与显著目标的整体结构。当背景在整幅图像中占据主导位置后,会导致前景Loss被稀释,在计算损失时会同等对待所有像素,忽略了像素间的差异,这些缺点会导致边缘模糊的问题,越靠近显著物体边缘的像素点,预测的准确率越低。本文依据标签中不同像素之间的差异,基于交叉熵和交并比两种函数,采用加权的方法区别对待不同像素之间的损失,可提高难样本点的预测精度,进而引导网络重点关注显著物体的全局和边界特征。6个常用数据集上的实验结果表明,相比近年的主流显著性检测方法,准确性有所提升。

In the field of salient object detection, the binary cross-entropy function is often used to calculate the loss. But this loss function calculates the information of each pixel point independently, ignoring the connection between each pixel and the overall structure of the salient object. When the background dominates the whole image, it will cause the foreground loss to be diluted. In addition, all pixels will be treated equally in calculating the loss, and the differences between pixels are ignored. The above disadvantages can lead to the problem of blurred edges, while the closer the pixel points are to the edges of significant objects, the lower the accuracy of prediction is. Based on the differences between different pixels in the labels, the weighted approach to treat the losses is used, between different pixels differently based on two functions of cross-entropy and cross-merge ratio, to improve the prediction accuracy of difficult sample points, and then guide the network to focus on the global and boundary features of the salient objects. The experimental results on six commonly used datasets show that the accuracy is improved compared to the mainstream saliency detection methods in recent years.