渐进聚合多尺度场景上下文特征的伪装物体检测

Progressively Aggregating Multi-Scale Scene Context Features for Camouflaged Object Detection

伪装物体检测通过模仿人类的视觉检测机理,实现在复杂场景下对伪装物体的定位与识别.然而,多数伪装物体检测方法在遇到相似外形目标干扰时,仅通过目标的局部表观特征无法准确识别伪装目标.为此,本文提出一种渐进聚合多尺度场景上下文特征的伪装物体检测网络,通过聚合多阶段语义增强的场景上下文特征来实现准确的伪装物体判别.具体来说,所提网络主要包含两个创新设计:U型上下文感知模块和跨级特征聚合模块.前者旨在感知复杂场景中物体的细节轮廓、纹理特征和颜色变化等丰富的局部-全局场景上下文信息.后者则结合坐标方向的注意力和多层级残差渐进特征聚合机制,逐级渐进聚合相邻层级之间的互补特征,实现对伪装物体全局语义的强化和局部细节的补充.本文方法在CHAMELEON、CAMO-Test、COD10K-Test和NC4K等4个非常具有挑战性的基准数据集上进行了评测.评测结果表明,本文方法相比于最新方法达到了领先的性能.

Camouflaged object detection(COD)is a computer vision task that imitates human visual mechanisms to recognize and locate camouflaged objects in complex scenes.However,the current COD methods cannot accurately discriminate the camouflage objects only by the local appearance features of the objects when meeting distractors with similar appearances.To this end,this paper presents a COD network based on progressively aggregating multi-scale scene context features,so that the accurate camouflaged object discrimination is realized by aggregating multi-stage semantic enhanced scene context features.Specifically,the network mainly has two novel designs:U-shape Context-Aware Module(UCAM)and Cross-level Feature Aggregation Module(CFAM).The UCAM aims to sense rich local to global context information such as detailed boundaries,texture features,and color changes of camouflaged objects.The CFAM combines the coordinate direction attention and the multi-level residual progressive feature aggregation mechanism to gradually aggregate complementary features between adjacent levels,strengthen the global semantics of camouflage objects and supplement local details.Extensive evaluations on 4 extremely challenging benchmarks including CHAMELEON,CAMO-Test,COD10K-Test,and NC4K,the experimental results have demonstrated that our model has achieved leading performance compared with state-of-the-art methods.