首先，现有的表征方法在性能上仍有很大的可提升空间，且缺乏统一的理论解释。本项目拟借鉴灵长类动物视网膜对客观世界的视觉表征和小波对信号的完全重建理论，构建描述子设计的统一框架，该框架既能统一现在主流的描述子设计方法，又能设计出性能更高的描述子。其次，现在图像表征和深度学习神经网络被视为两个互不联系的问题割裂开来研究，且后者消耗大量的样本和计算。而从生物视觉系统的视觉通路来看，视网膜对客观世界的完备表征为大脑神经网络不变特征的提取提供了有效的基础，且视网膜的表征是固定的不参与大脑皮层的学习过程。 基于此，本项目拟基于对人类视觉通路的认知构建将图像表征、深度神经网络结合在一起的统一方法，通过固定神经网络前面层的处理，减少深度神经网络对训练样本的依赖。最后，生物神经元之间的通讯是通过发放电脉冲，本项目拟基于生物脉冲的编码机制，将上述所得神经网络转化为脉冲神经网络，进一步提高计算效率。

First, existing image representation methods not only have large development room in performance, but also are failed to explain themselves thanks to the absence of unified theory. Our project plans to build a unified frame to explain the mechanism of representation methods and instruct the designation of descriptors, based on the reference to visual representation of primate retina to objective world and complete signal reconstruction theory in wavelet analysis. This frame can unify popular descriptors as well as develop new ones with high performance. Second, image representation method and deep neural network (DNN) are considered as different study tasks with no direct relation. Furthermore, DNN requires a mass of training samples and computation resources. In biological visual pathway, however, the full representation of retina to visual world provides a basis for subsequent central neural system to extract invariant features. Moreover, the structure of retina is fixed, without participating in the learning process of cortex. Based on the knowledge of human visual pathway, our project intends to develop a method combining the representation and deep neural network approaches. Therefore, by fixing the earlier layers in DNN, the dependencies of sample and resources could be reduced. Last but not least, neurons in biological neural system communicate by spikes. In order to further improve the computational efficiency, we turn the Artificial DNN into Spiking Neural Network on the strength of the time coding mechanism in the spiking trains of real neuron.