基于兴趣区域掩码卷积神经网络的红外-可见光图像融合与目标识别算法研究

Research on Infrared Visible Image Fusion and Target Recognition Algorithm Based on Region of Interest Mask Convolution Neural Network

建立权重独立的双通道残差卷积神经网络,对可见光与红外频段下的目标图像进行特征提取,生成多尺度复合频段特征图组。基于像点间的欧式距离计算双频段特征图显著性,根据目标在不同成像频段下的特征贡献值进行自适应融合。通过热源能量池化核与视觉注意力机制,分别生成目标在双频段下的兴趣区域逻辑掩码并叠加在融合图像上,凸显目标特征并抑制非目标区域信息。以端到端识别网络作为基础,利用交叉损失计算策略,对含有注意力掩码的多尺度双频段融合特征图进行目标识别。结果表明,所设计的识别网络能够有效地融合目标红外热源物理特征和可见光图像纹理特征,提高了信息融合深度,保留目标热辐射与纹理特征的同时降低了背景信息干扰,对全天候复杂环境下的多尺度热源目标具有良好的识别精度与鲁棒性。

A dual channel residual convolution neural network with independent weight is established.The features of target in visible and infrared images is extracted.The multi-scale composite frequency band feature maps are generated.Based on the Euclidean distance between image points,the saliency of each image point in the dual band feature map is calculated.The adaptive fusion is carried out according to the characteristic contribution value of the target in different imaging frequency bands.Through the thermal radiation pooling kernel and visual attention mechanism,the logical mask of the target region of interest under dual frequency band is generated and superimposed on the fusion image to highlight the target features and suppress the non target area.Based on end-to-end identification network and using the cross loss calculation strategy. The target recognition of multi-scale dual band fusion feature map with attentionmask is carried out. The results show that the designed recognition network can effectively integrate thephysical characteristics of infrared heat source and the line features of visible image. The depth ofinformation fusion is improved. The thermal radiation and texture features of the target is retained. Theinterference of background information is reduced. It has good recognition accuracy and robustness formulti-size heat source targets in all-weather and complex environment.