迷彩伪装目标检测的视觉特征偏好研究

Research on Visual Feature Bias of Camouflaged Object Detection

迷彩伪装通过设计颜色和纹理图案来破坏目标的固有形状,其检测依赖的视觉特征应与常规目标不同。然而卷积神经网络的黑盒性质使得不同视觉特征对模型识别的贡献程度无法获知。为解决上述问题,借鉴人类视觉系统设计了一种适用于伪装场景的视觉特征解耦方法,解耦并分析目标检测模型在颜色、纹理和形状特征上的偏好程度。具体来说,使用消除单一特征并保留其余特征的解耦框架,以模型的性能下降情况作为偏向性的衡量标准,通过灰度化处理消除图像的颜色特征,使用区域置乱破坏目标的纹理特征,对目标轮廓取内接形状以改变目标的形状特征。在公开的迷彩伪装人员数据集和常规人员检测数据集上分别进行实验,结果显示,迷彩伪装目标的检测主要依赖纹理,常规目标的检测主要依赖形状。

Camouflage uses designed color and texture patterns to disrupt the inherent shape of the target,so the visual features that its detection relies on should be different from those of conventional targets.However,the black box nature of convolutional neural networks makes it impossible to know the contribution of different visual features to model recognition.To solve this problem,a new visual feature decoupling method was designed based on the human visual system,which is suitable for camouflage scenes.This method decouples and analyzes the preference degree of object detection models on color,texture,and shape features.Specifically,an analysis architecture was used to eliminate a single feature while retaining the remaining features,and the performance degradation of the model was used as a measure of bias.Grayscale processing was used to eliminate the color features of images,region scrambling was used to disrupt the texture features of targets,and the inner shapes of targets were extracted to change their shape features.Experiments were conducted on publicly available datasets of camouflaged personnel and conventional personnel detection,respectively,and the results showed that the detection of camouflaged object mainly relies on texture,while the detection of conventional object mainly relies on shape.