A feature binding model in computer vision for object detection

In this paper, the authors propose the “Feature Binding (FB)” strategy in computer vision, a method combined with the biological visual perception theory. Based on feature subspace, the proposed method refers to the biological model and binds features according to certain rules. All features bound in a group are taken as a whole. Besides, all groups with different weight coefficients according to different importance are used to determine the object and its location. The position of the object can be determined based on the calculation according to the corresponding criteria. Feature Binding can significantly enhance the accuracy of object detection and localization. Moreover, the method can accelerate object detection and resist external interference in the unbound feature subspace. Feature Binding has good accuracy not only for the whole object but also for the obscured object. It also has good robustness for different algorithms, which are based on features, including traditional methods and deep learning algorithms. The object positioning system can detect the partially occluded objects more accurately in practice.