Adaptive Intelligence for Robust Image Analysis
The automated interpretation of images to detect and recognize objects in a timely manner is crucial in vision tasks, especially in applications related to surveillance and monitoring, autonomous navigation, and industrial robotics, among others. Over the years, a multitude of approaches and algorithms have been developed. However, most approaches are developed for a specific application and cannot be generalized for all images. In fact, no single algorithm can be considered good for all images, nor are all algorithms good for a particular image. Each algorithm's utility is limited by its specific characteristics that makes it applicable for particular kind of images. The fundamental challenge in building intelligent systems is then to provide a generalized framework that is capable of choosing a suitable algorithm from many candidates given a particular image.
Performance Prediction Framework
Image analysis is basically a problem of psycho-physical perception, and therefore not susceptible to a purely analytical solution. Based on the knowledge of an algorithm's characteristics, humans are able to predict the performance of every algorithm on a given images, and thus choose an optimal one. Simulating this process can lead to automation in predicting the performance of various algorithms and hence the eventual selection of an optimal one given the input image. Inherently, the performance of an algorithm can include a variety of measures such as processing quality, stability, time and memory complexity, etc. To incorporate image context information and knowledge about an algorithm's performance, we are developing a performance prediction system that comprises of three main modules: feature extractor, performance evaluator, and a predictor. The feature extractor is meant to identify the context of the image. The performance evaluator simulates the human observer. Finally, the predictor provides a convenient mechanism to automatically acquire, store, and utilize human knowledge in an implicit way. Specifically, the predictor combines information from the feature extractor and the performance evaluator, which forms the basis of the image context and the knowledge about each algorithm's performance characteristics. The predictor uses this information to estimate the outcome of each candidate algorithm. The one predicted to generate the most desirable outcome can then be chosen. A general schematic of the system is shown heer.
