Fuzzy image correlator with existence measure for fuzzified images

Qinan Mao, Purdue University

Abstract

Object recognition and localization are often difficult by stereo images; partly because the image space is discrete and partly because the intensity of the images is corrupted with unknown noise, intensity quantization, distortion of geometrical mapping, and so on. This research deals with three important problems based on a lateral stereo model. The first topic is the object recognition and localization on a single fuzzified image. First, the uncertain data, often unable to be modeled by traditional methods, are identified and represented by compact fuzzy intervals. Then, a fuzzy image correlator (FIC) is proposed to perform the object recognition and localization. This FIC is capable of tolerating object distortion and aggregating the uncertain information to the output, represented by a possibility distribution of the position of the desired object. A simplified fuzzy image correlator (SFIC) is also devised by only choosing the significant parts of the reference object for the correlator design. The proposed FIC/SFIC can also be applied to partial object recognition. The performance of the proposed FIC/SFIC is verified by conducting experiments with distinct objects and partially occluded objects embedded in an arbitrary background. The second topic is the 3-D object recognition and localization by extending the lateral stereo model and the FIC to a pair of stereo images. Based on the proposed FIC, the desired object is located on the stereo images independently. Then, their possibility distributions are propagated to depth planes by a fuzzy relation. The desired object can be located by examining the possibilities on each depth plane. Again, experimental study was performed to verify the performance of the proposed FIC. The third topic is the surface reconstruction on smooth surface. This problem is often overlooked and is solved by an interpolation method. We intend to apply the proposed FIC/SFIC to the recognition of objects with smooth surfaces in range-based images. We reformulate it to searching corresponding scale-invariant features extracted from the smooth surfaces in individual epipolar lines. Then, two Hopfield neural networks are utilized independently to reconstruct the smooth surfaces, based upon the smoothness of depth and the intensity of each image. The extension of the FIC/SFIC to recognizing objects with smooth surfaces in range-based images is an item in our future research.

Degree

Ph.D.

Advisors

Lee, Purdue University.

Subject Area

Electrical engineering

COinS