基于渐进式嵌套特征的融合网络

显著目标检测是指通过引入人类视觉注意力机制,使计算机能检测视觉场景中人们最感兴趣的区域或对象.针对显著性目标检测中存在检测边缘不清晰、检测目标不完整及小目标漏检的问题,文中提出基于渐进式嵌套特征的融合网络.网络采用渐进式压缩模块,将较深层特征不断向下传递融合,在降低模型参数量的同时也充分利用高级语义信息.先设计加权特征融合模块,将编码器的多尺度特征聚合成可访问高级信息和低级信息的特征图.再将聚合的特征分配到其它层,充分获取图像上下文信息及关注图像中的小目标对象.同时引入非对称卷积模块,进一步提高检测准确性.在6个公开数据集上的实验表明文中网络取得较优的检测效果.

Fusion Network Based on Progressive Nested Feature

In salient object detection, the computer detects the most interesting areas or objects in the visual scene by means of introducing the human visual attention mechanism. Aiming at the problems of unclear edge, incomplete object and missing detection of small objects in salient object detection, a fusion network based on progressive nested feature is proposed. Progressive compression module is adopted to continuously transfer and merge deeper features downward and make full use of advanced semantic information while the number of model parameters is reduced. A weighted feature fusion module is designed to aggregate the multi-scale features of the encoder into a feature map that can access both high-level and low-level information. Then, the aggregated features are allocated to other layers to fully obtain image context information and focus on small objects in the image. The asymmetric convolution block is introduced to further improve the detection accuracy. Experiments on six open datasets show that the proposed network achieves good detection results.