Incremental feedback learning mechanism for highly efficient automatic image segmentation with feature coupling

Segmentation of images is a pre-requisite for all modern high efficiency image processing system. In order to perform this task, various application specific algorithms are designed and deployed by image processing experts. These systems work on a context-specific mode, wherein all segmentation outputs are restricted by context of images for which the system is trained. In order to deploy these systems to other domains, complex tuning and optimization operations are needed. This reduces applicability of these system models for real time use cases, where general purpose segmentation methods are needed. These use cases include but are not limited to, scene classification, satellite image classification, yield prediction, traffic detection, etc. Moreover, general purpose image segmentation models work effectively only under a pre-set types of application scenarios, and need to be constantly trained in order to improve their applicability. Retraining these systems increases computational costs, and requires large training and testing delays. In order to remove these drawbacks, in this text an incremental feedback learning mechanism with feature coupling is proposed. The proposed model uses a wide variety of image segmentation methods that analyze colour, texture & shape information; and map it with relevant image features. These features are traced along with segmentation quality metrics like peak signal to noise ratio (PSNR), figure of merit (FOM), minimum mean squared error (MMSE), and probabilistic random index (PRI) in order to evaluate the best segmentation algorithm. These features are classified using an ensemble classification model for selection of the most efficient segmentation method that maximizes PSNR, & PRI while minimizing MMSE. Parametric evaluation suggests that the proposed model is able to improve segmentation accuracy by 8%, and reduce false alarm rate by 15% when compared with standard automatic segmentation models.