增强特征金字塔结构的显著目标检测算法

Salient Object Detection for Enhanced Feature Pyramid Structure

显著目标检测是计算机视觉的研究热点。显著目标检测算法存在一些问题,如:算法常采用单一损失函数,缺乏对多维特征损失的考虑,可能带来局限性;最高层特征图来源单一;特征图融合常使用对应像素相加,不能有效突出图像中感兴趣区域。针对上述问题,结合结构性相似、交并比和交叉熵三种损失函数来捕捉图像细节,采用对应像素相乘操作融合特征图,令模型对显著区域更加敏感;通过残差特征图增强模块逆向构建更高层特征图强化其语义信息;采用特征金字塔结构融合不同尺度信息,完成编码解码模块。在5个数据集的对比实验表明该方法性能超过主流算法,能实现有效的显著目标检测。

Salient object detection plays an important role in the field of computer vision and is a hot research spot.However,there are some problems in its algorithm,such as the single loss function often ignores the influence of multi-dimensional features;the fusion of feature images usually depends on simple addition of pixels,it cannot effectively highlight the region of salient object in the image.In order to overcome above problems,this paper captures the image details by combining structural similarity,intersection over union and cross entropy loss functions.Fusion of feature maps is with multiplication operation that can make the model more sensitive to the salient region.Residual feature map enhancement module is used to build the higher level feature map to strengthen its semantic information.Finally,the feature pyramid structure is used to fuse the information of different scales to complete the coding and decoding module.Experimental comparison on five benchmarks show that the performance of the presented method is better than the popular methods and can detect salient object effectively.