2.  ImageFinder Overview
2.1   ImageFinder Internal Structures
2.2   Image Preprocessing
2.3   Normalization
2.4   Feature Recognition
2.5   ABM Matching Engine
2.6   Multi-Layered ABM
2.7  Chapter Overview

2.  ImageFinder Overview

2.1   ImageFinder Internal Structures

ImageFinder adopts a layered approach to image recognition. An image recognition application is roughly divided into:

Level 5: User Interface;
Level 4: Data Management;
Level 3: Image-Matching Layer;
Level 2: Scanner, Camera, and their Drivers;
Level 1: PC with Windows.
The reason for the above approach is that the ImageFinder library will be integrated into a database application. Stand-alone software is roughly divided into 3 layers:
Level 3: User Interface;
Level 2: Image-Matching Layer;
Level 1: PC with Windows.
For the ImageFinder, the Image-Matching Layer is divided into:
Image Preprocessing
Normalization
Feature Recognition
Pixel Recognition
Sub-Pixel Recognition
Each of the sub-layers is further divided into filters:

Image Preprocessing

Edge Filters;
Threshold Filters; and
Clean Up Filters.
Normalization
 Reduction Filter.
Feature Recognition
BioFilter;
NeuralFilter.
Pixel Level Recognition
 ABM Filter.
Multi-Layered Pixel Recognition
BioFilter 2;
NeuralFilter 2;
ABM Filter 2.
�

The ImageFinder uses the following GUI (Graphical User Interface) rules:

An image will flow through the ImageFinder via the above 10 filters; however, in a typical application, the image can also bypass many of the filters.

The combination of the 10 filters will have thousands of free parameters to adjust. The ImageFinder has only about 70 parameters available for adjusting by users.

2.2   Image Preprocessing

The image preprocessing sub-layer prepares the image for the ImageFinder. The image pre-processing process is not unique; there are many options available. Some are better than others. You should choose an image pre-processing procedure where the sample objects stand out, otherwise change the options. If you do not have a good pre-processing setting in the off-the-shelf ImageFinder, a customized filter has to be built. Do not make too many things stand out, i.e. as long as the area of interest stands out, the rest should show as little as possible.

Image pre-processing can determine the Recognition Rates. For many problems like fingerprints, palm prints, �, a special image processing procedure will be required.

The Image Pre-Processing Layer consists of three types of filters:

Edge Filters (Optional);
Threshold Filters (Required); and
Clean Up Filters (Optional).
The ImageFinder applies these three filters in the above order.

The Edge Filters attempt to exaggerate the main features a user is looking for.
The Threshold Filters attempt to suppress the background.
The Clean-Up Filters will smooth the resulting image to reduce recognition error.

If you are not familiar with image processing, please try the following:

Setting 1:

Edge Filters: Sobel 1 (or Sobel 2)
Threshold Filters: Dark Background 128
Clean Up Filter: Medium.
Setting 2:
Edge Filters: None (�Enter An Edge Filter�)
Threshold Filters: Light Background 128
Clean Up Filter: None (�Enter A CleanUp Filter�)
Beyond that, please use the trial and error approach to exaggerate the main features you are looking for and to suppress the background.

2.3   Normalization

The Normalization sub-layer will prepare the images for the underlying Image Matching Engine. The Attrasoft Image Matching Engine is an internally developed algorithm, which is called �Attrasoft Boltzmann Machine� or ABM. The ABM neural net deployed in the ImageFinder, by default, is a 100x100 array of neurons.

While any size of ABM neural net can be used, when coming to a particular application, a decision has to be made. The ImageFinder uses 6 different sizes:

Later in the multi-layered design, the number of neurons can be much larger. The Reduction Filter will connect the images to various sets of ABM neural networks.

2.4   Feature Recognition

Before the ABM Engine processes an image, it will go through a pre-matching step called Feature Space Matching or Feature Matching. The purpose of Feature Matching is to eliminate unmatched images. This Feature Matching sub-layer has two filters:

BioFilter; and
NeuralFilter.
A typical image has a dimension of say 480x640, or 300,000 pixels. A Feature Space of an image is a collection of variables computed from a given image. The number of variables in a Feature Space is usually less than 1% of the number of pixels. As a result, the matching in the Feature Space is much faster than pixel matching.

For example, after FFT, (Fast Fourier Transform), the first few terms will be variables in a Feature Space. (The Fourier Transform is a way to convert Space-domain data into its frequency components).

The BioFilter will attempt to eliminate 80% of the mismatches, and the Neural Filter will attempt to eliminate 19% more of the mismatches, leaving only 1% for the slower matching.

2.5   ABM Matching Engine

The ABM neural network is the matching engine of the ImageFinder. There is one filter in this sub-layer, the NeuralNet filter or ABM Filter. The ABM Matching Engine is responsible for all of the Attrasoft products. The ABM has been applied to many fields of pattern and image recognition. Each time a customer is served, their feedback polishes this core technology. This core technology has been tested over and over again since 1995.

2.6   Multi-Layered ABM

Once we go beyond one layer of the neural network, there are endless choices. An application will normally dictate the selection of a Multi-Layer ABM Architecture. In this version, an example of Multi-Layered ABM Architecture will be used, which is based on the ImageExaminer. Depending on the arrangement, this sub-layer can have many filters. In this version, this sub-layer has 3 additional filters:

BioFilter 2
NeuralFilter 2
ABM Filter 2


2.7   Chapter Overview

Part I. Customized ImageFinder

In this version, the number of parameters is increased; the number of filters is increased; and the number of commands is increased. As a result, this version is significantly harder to use than the earlier version. For this reason, we will present three customized examples first.

Customized Versions are easy to use, usually a few clicks, so you can see once the tuning is over, the operation of the software is very easy.

Chapter 3, FaceFinder, presents a customized example for the front view facial images in the Feret database.

Chapter 4, DocumentFinder, presents a customized example for scanned document images.

Chapter 5, ImageExaminer, presents a customized example for comparing two basically identical images and identifying some minor differences. There are three chapters later that are related to this customized software, BioFilter II, Neural Filter II, and ABM Filter II.

Part II. Basic Architecture.

The Basic Architecture is a one-layer neural-network architecture:

Image Preprocessing

Edge Filters;
Threshold Filters; and
Clean Up Filters.
Normalization
 Reduction Filter.
Feature Recognition
BioFilter;
NeuralFilter.
Pixel Level Recognition
 ABM Filter.
The ImageFinder recognizes images in two phases: The purpose of Feature Space Matching is to eliminate unmatched images. This Feature Recognition sub-layer has two filters:
BioFilter; and
NeuralFilter.
The main image recognition uses the NeuralNet Filter.

Chapter 6, Image Processing, briefly describes the image processing process required for the ImageFinder. You have to set three filters, Edge Filters, Threshold Filters, and Clean Up Filters. The Threshold Filter is required; the other two filters are optional.

Chapter 7, BioFilter, will describe the minimum number of steps for using the BioFilter for image recognition:

Initialization sets the ImageFinder parameters. Then, the image signatures are calculated and stored in a record. Training teaches the BioFilter who matches with whom. After training, the BioFilter can be used for 1:1 and 1:N Matching.

Chapter 8, NeuralFilter, introduces the Neural Filter. The Neural Filter is the second classification filter in the feature space; it runs parallel to the BioFilter.

Chapter 9, NeuralNet Filter or ABM Filter, introduces the neural network used in the ImageFinder. The NeuralNet Filter commands are divided into three types: Short, Long, and File depending on the input. Short-Search processes images in one directory, Long-Search processes images in sub-directories and File-Search processes images in input files. The File-Search will use the output of the Feature Space recognition as input.

Chapter 10, Parameters, gives a more detailed description of all parameters used in various filters.

Chapter 11, Batch, introduces the batch command, which allows you to save your setting and execute your problem in a few clicks. The batch file can also be used in the dos version.

Part III. More on NeuralNet Filters

Chapter 12, ABM Filter Short-Search, introduces Short-Search, which will search images in a directory.

Chapter 13, ABM Filter Long-Search, introduces Long-Search, which will search images in sub-directories.

Chapter 14, ABM Filter Parameter Advisor, introduces �Parameter Advisor�, which will help you to select the neural net parameters.

Chapter 15, Biometrics, introduces a Fingerprint Recognition Example.

Chapter 16, Segment-Locator, introduces several Commands, and Examples that will locate a segment in images.

Chapter 17, Point-Locator, introduces several Commands, and Examples that will locate points in images.

Part IV.  Extended Architecture

When you go beyond the Basic Architecture, the options are limitless.
We will focus our discussion on a specific application: to compare two basically identical images and identify the minor differences.

Our approach is to divide the image into smaller sections and use our Basic Architecture for each smaller section of the image. This will generate the following additional layers:

BioFilter II;
NeuralFilter II;
ABM Filter II.
Chapter 18, BioFilter II, introduces BioFilter II, which is similar to the BioFilter but operates on an image segment rather than the whole image.

Chapter 19, Neural Filter II, introduces Neural Filter II, which is similar to the Neural Filter but operates on an image segment rather than the whole image.

Chapter 20, ABM Filter II, introduces ABM Filter II, which is similar to the ABM Filter but operates on an image segment rather than the whole image.

Part V.  Developers

Software developers can integrate the ImageFinder technology into their applications. The ImageFinder Dos Version is a quick and easy way.

Chapter 21, ImageFinder Dos Version, introduces a quick way to implement an image recognition application.

Chapter 22, API, introduces the standard Application Programming Interface (API) of the ImageFinder for software developers. The Image Recognition Level Library is called TransApplet.

Chapter 23, Reference Manual, lists and explains each of the Filters, Commands, and Parameters.

Chapter 24, ImageFinder Support, lists the support packages.

Chapter 25, Readme.txt, lists the contents of the readme.txt file.
 

Return