Image Segmentation By A Simplified PCNN With Automatic Parameter Setting Method And Its Application To Object Recognition

In this thesis, we propose an automatic parameter setting method of a simplified pulse coupled neural network (SPCNN) for image segmentation. Our method successfully determines all the adjustable parameters in SPCNN and does not need any training and trials as required in previous studies. In order to achieve this goal, we derive the general formulae of dynamic threshold and internal activity of the SPCNN according to the dynamic properties of neurons, and deduce the sub-intensity range expression of each segment based on the general formulae. Furthermore, we extract information from an input image, such as the standard deviation and the optimal histogram threshold of the image, and build a direct relation between the dynamic properties of neurons and the static properties of each input image. Finally, the experimental segmentation results of the gray natural images from the Berkeley Segmentation Dataset, rather than synthetic images, prove the validity and efficiency of our proposed automatic parameter setting method of SPCNN.We also propose a view-based object recognition method to identify objects from complex real-world scenes, named "SPCNN Region-based Matching Object Recognition", i.e., SPCNN-RMOR. The proposed SPCNN-RMOR method employs the simplified pulse coupled neural network (SPCNN) to segment color model object image and test image, and further performs region based matching between the model object image and test image. Hereinto, the parameters of SPCNN are automatically set by the aforementioned method in terms of each model object. And a color transformation achieving invariance of light intensity change (scale-invariant) and invariance of light intensity shift (shift-invariant) is performed on both images before the SPCNN process. Furthermore, a novel image segmentation strategy is introduced to exploit both temporal information of SPCNN pulses and spatial distributions of images (while the previous PCNN feature extraction approaches usually lose the spatial information by encoding the spatial distributions of a2-D image into a1-D temporal sequence), which leads to fine enough image segmentation results. And to a large extent, the image segmentation process under the novel image segmentation strategy helps to overcome the drawback of feature-based object recognition methods that inevitably include background information into local invariant feature descriptors when keypoints locate near object boundaries. In addition, certain adaptive thresholds which are adjustable according to the specific model object are employed in stages of outlier region blobs removal, cluster formation and clusters refinement. As a result, the test cluster with the minimal Bhattacharyya distance of color histograms has the highest probability of containing the desired object. Finally, a large number of object recognition experiments prove that the proposed SPCNN-RMOR object recognition method has strong robustness against the variations in translation, rotation, scale and illumination, and affine/perspective distortions, under partial occlusion and highly clutter backgrounds. Moreover, it shows a good performance in identifying objects with less texture, which significantly outperforms most mainstream feature-based object recognition methods.In summary, the proposed automatic parameter setting method of SPCNN and its application to color image segmentation have laid a foundation for the proposed SPCNN-RMOR object recognition method which could identify objects from complex real-world scenes.