An Introduction to Artificial Neural Networks for Image Processing

Artificial Neural Networks (ANN) are a type of computer programming that is designed to mimic the process of human cognition found the natural neural network – the human brain. Artificial Neural Networks have been successfully applied to tasks involving image compression, classification, pattern matching, and change detection. This presentation summarizes the current research of the author related neural networks for image processing. These interests include extracting feature primitives from raster images and LiDAR, the chunking of data into larger groups of information, and ultimately change detection. Introduction Artificial neural networks represent a very different form of computer programming. Neural networks are an unstructured approach to programming unlike the structured programming environments of C++, Visual Basic, *.NET, etc. The unstructured approach makes them very interesting tools for performing a variety of complex logic operations, especially those operations that may not be reducible to a series of rules or steps that can be sequentially coded in a serial, logical method. Neural networks are good for solving unstructured problems, making quick approximations, and storing data. They are suited for pattern recognition in both signals and pictures. They excel with issues of change detection and data optimization problems like the traveling salesman and the Chinese postman. A variety of other applications exist for neural networks that offers a wide range of opportunities for remote sensing and geographic information systems. A special ability of the neural network includes their ability to learn. By teaching a neural network what outputs are expected from many and varied inputs, their programmer is actually their teacher. Structure & Terminology Neural networks are designed to emulate the human cognitive process. As a computation model, neural networks consist of many interconnected simple processors, similar to the neurons in the human brain. These neurons are capable of performing threshold logic operations like yes or no. Input/output functions can be performed via the internal arrangement of the connections between the neurons, which are termed “synapses.” Neural networks are often described as operating in parallel or in a continuous fashion, because they are capable of accepting many parallel inputs, processing the inputs in parallel, and outputting continuous results in near real-time. However, most neural networks are implemented on serial computers (Intel, Motorola, etc.), so they only simulate parallel processing.