基于边缘与注意力的递进式伪装目标检测

Progressive Camouflage Object Detection Based on Edge and Attention

摘要:

本文针对现有的伪装目标检测(Camouflage Object Detection,COD)网络模型不能充分利用各个层次之间的信息从而导致检测结果模糊的问题,提出了基于边缘与注意力的递进式网络(Progressive network based on edge and attention,EAPNet)。首先提出了边缘信息提取模块(Edge Information Extraction Module, EIEM),提取主干网络第一、三、五阶段的特征生成边缘特征图；然后提出了感受野与边缘特征聚合模块(Receptive Field and Edge Aggregation Module, RFEAM)聚合主干网络相邻阶段的感受野特征和边缘特征；最后利用所提出的注意力辅助融合模块(Attention-Assisted Fusion Module,AAFM)对相邻RFEAM的输出特征进行二次递进式融合。在三个COD数据集上的实验结果表明本文方法优于其他10个模型。

Abstract:

In this paper, a progressive network based on edge and attention (EAPNet) is proposed to address the issue of insufficient inter-layer information utilization in existing Camouflage Object Detection (COD) models, which leads to blurred detection results. First, an edge information extraction module (EIEM) is proposed to extract the features from the first, third, and fifth stages of the backbone network to generate edge feature maps; next, a receptive field and edge aggregation module (RFEAM) is proposed to aggregate receptive field features and edge features from adjacent stages of the backbone network; finally, the proposed attention-assisted fusion module (AAFM) is used to perform secondary progressive fusion on the output features of adjacent RFEAMs. Experimental results on three COD datasets show that the proposed method outperforms ten other models.