Object detection based on saturation of visual perception
Object detection based on saturation of visual perception
复制标题
DOI：
10.1007/s11042-020-08866-x
复制
发表时间：
2020-04-03
影响因子：
3.6
通讯作者：
Yan, Wei Qi
中科院分区：
文献类型：
--
作者：
Pan, Chen;Yan, Wei Qi
本文提出了一个框架，通过模拟人类的感知，从自然图像中检测显著对象。我们认为，由凝视中的微扫视产生的感知饱和是人类大脑输出显著物体意识的主要原因。感知饱和度是用微跳幅度（AOM）表示的。当AOM或AOM的变化趋于零时，人类感知变得饱和。出于这种分析，我们构建了一组基于学习的模型来检测一个显着的对象在一个由粗到细的序列。首先，选择一个小的图像，以最小化AOM，从而饱和的感知。然后，随机权值神经网络（NNRW）被用来模拟接收到的视觉刺激。为了检测AOM的变化，构造了一个正反馈回路，迭代地执行“像素采样-学习分类”过程。迭代后的最终注视区域被视为显著对象。该算法完全基于无监督学习和数据驱动。基于开放图像数据集的实验结果表明，与现有的无监督算法相比，该方法具有更好的性能。
This paper presents a framework to detect salient objects from natural images through simulating human perception. We think perception saturation, generated by microsaccades in gaze, is the primary reason why human brains export consciousness of salient objects. Perception saturation is represented by using amplitude of microsaccades (AOM). When the AOM or changes of the AOM tend to zero, human perception becomes saturated. Motivated by this analysis, we construct a group of learning-based models to detect a salient object in a coarse-to-fine sequence. Firstly, a small image is selected to minimize the AOM so as to saturate the perception. Then, neural networks with random weights (NNRW) are chosen to simulate the received visual stimuli. In order to examine the changes of AOM, a positive feedback loop is constructed which executes the procedure of "pixel sampling-learning classification" iteratively. The final fixation area after iterations is regarded as a salient object. The proposed algorithm is based on unsupervised learning and data-driven completely. Our results based on open image datasets show that the proposed method achieves better performance compared to those existing unsupervised algorithms.