基于视觉注意的目标预检测模型

Target Pre-detection Model Based on Visual Attention

提出了一种从输入图像中提取一致性特征并将该特征融入视觉注意模型中的方法,它以Itti提出的视觉注意模型为基础,首先采用线性离散高斯滤波器对输入图像进行平滑和降采样,得到不同分辨率的子图像,建立输入图像的高斯金字塔结构。然后利用梯度结构张量计算各子图像的一致性并通过中央周边差操作得到一致性特征图,最后利用层间求和操作求得输入图像的一致性显著图引导注意。实验表明,在视觉注意模型的初级视觉特征提取阶段添加图像的一致性特征,能检测出Itti模型不能检测的一致性特征显著的感兴趣区域,而且能优先检测出具有特殊方向的目标。

A method for computing the coherence feature of the input image and adding it to the visual attention model is developed based on the visual attention model proposed by Laurent Itti. The method uses a linearly separable Gaussian filter to smooth and sub-sample the input im- age into maps with different resolution levels to obtain the pyramidal representation. Then, a gradient structure tensor is used to obtain the coherence features of each map at different scales. After that, center-surround receptive fields are simulated by across-scale subtraction yielding ＂coherence feature maps＂. Finally, the coherence feature map is summed up by using across-scale addition to form a ＂coherence conspicuity map＂ to guide visual attention. Experiments on different images show that after adding coherence feature into the Itti model, regions of interest with the strong coherence feature, which fails to be extracted by the original model, and targets with special orientation can be detected with high priority.