Research On Biologically Inspired Object Recognition Algorithms

Biological visual system is the most powerful and comprehensive visual system. Its structural and related mechanism could inspire the design of new computer vision sys-tems. In recent decades, neural science has provided us deeper insights over biologi-cal visual system. Useful guidelines could be concluded from biological mechanisms. They could be applied to the design and implementation of object recognition algo-rithms.From the perspective of the overall design of the computer vision system, we an-alyze some instructive biological mechanisms and discuss their relationship with com-puter models. Based on the analysis and discussion of biological mechanisms, some guiding principles of the model design are summarized, including hierarchical struc-ture, receptive field, feature learning, computation locality, structural consistency, sig-nal feedback, and attention selection. The advantages and disadvantages of some object recognition algorithms are analyzed using these principles. Then these guiding princi-ples are applied to the actual design of several computer vision models:1. Utilizing biologically inspired principles on hierarchical structure, feature learn-ing, computation locality, structural consistency, and signal feedback, we pro-pose a multilayer in-place learning network, which is built upon a unified neuron model. The model integrates unsupervised learning and supervised learning un-der the same framework. The supervising signal is back propagated through soft supervision mechanism, while avoiding problems like local minima and vanish-ing gradient. Through learning and adaptation to the environment and the task, the model incrementally develops a hierarchy of internal representations with in-creasing invariance from early layers to later layers. Soft supervision improves the type purity of the neurons, which could help the generation of better features and improve the recognition rate. The hierarchical internal representations could be shared among different tasks, which could accelerate the training speed of new tasks.2. Based on the multilayer in-place learning network, we model the functional layers of the cortex. Experiments show that modeling different functions using different layers is necessary. We also model the adaptive lateral connections based on functional layers. After training over Gaussian spots and gratings, the neurons generate lateral connection weights in the shape of Mexican-hat. This could help the design and implementation of more biologically plausible models.3. Utilizing biologically inspired principles on hierarchical structure, receptive field, feature learning, structural consistency, and signal feedback, we discuss the opti-mal low level feature learning method in a multilayer neural network. We com-pare different ways of selecting receptive fields in a multilayer neural network through an object recognition task. In the experiment, the model based on the the biologically inspired design principles achieved the highest recognition rate. The receptive fields of low level neurons take the shape similar to simple cells. This validates two biologically inspired design guidelines. The receptive fields should be learned through training samples so that they could decorrelate the input and reduce redundancy. And during the training phase of the classification task, the receptive fields should adapt for the classification task.All related experiments show that biologically inspired design guidelines could help improve object recognition algorithms.