人类视觉注意机制在目标检测中的应用

Applications of human visual attention mechanisms in object detection

根据人类视觉感知理论,在介绍了两种比较有代表性的视觉注意模型的基础上,采用bottom up控制策略的预注意机制和top down控制策略的注意机制,提出了一种适用于自动目标识别的目标检测算法。从输入图像出发,采用Gabor算子建立多尺度、多方位的多通道图像,通过全波整流和各通道间的对比度增益控制,得到多尺度、多方位的方位特征图,这些特征图的线性组合则为显著性图。给出了仅采用bottom up控制策略的船舶目标检测实验结果,待检测目标在显著性图中得到明显增强,有利于检测的实现。

Two models of visual attention which are consistent with human visual perception are introduced. Based on these two models, an ATR algorithm employing attention mechanisms with bottom-up and top-down control strategies is developed. A multi-channel representation of the input image was obtained by a bank of Gabor filters, corresponding to multiple scales and multiple orientations. Full-wave rectification of each channel and contrast gain control among the channels were performed sequentially to get orientation feature maps with multiple scales and multiple orientations. The so-called salient map was the linear combination of these orientation feature maps. The experimental result with ship detection is given by use of bottom-up control strategy only. The objects for detection are effectively enhanced on the result map.