Towards incomplete object recognition

Human visual system can recognize incomplete contours and objects easily. However, these kind of recognition tasks are extremely challenging in computer and robot vision. This paper demonstrates how combination of genetic algorithms and morphology operations can be used to generate an image processing procedure for recognition of subjective objects (e.g. incomplete objects). In order to acquire the optimal object recognition procedure, the approach receives the subjective object and the corresponding user-prepared gold sample (physical object which reflects the user's visual expectations). After carrying out the training or optimization phase, the optimal procedure is generated and ready to be applied on new subjective objects (the same object but with different incomplete parts, sizes, etc.). No dependency on domain knowledge or huge number of sample images, one time training for a group of images, and robust object recognition are the novel features of the proposed approach. The training takes place based on one gold sample. This desirable characteristic reduces the level of dependency on expert participation which is usually an obstacle for full automation in most applications. The approach architecture and the employed methodologies are explained in detail. The performance of the approach has been evaluated by several well-prepared experiments to recognize various incomplete objects.