1/3/2014
CONTENTS:
IMAGE PROCESSING
1/3/2014
IMAGE PROCESSING
Image processing is the field of research concerned with the development of computer algorithms working on digitized images
1/3/2014
image processing problems are solved by a chain of tasks or the pipeline which consists of the steps of pre-processing, data reduction, segmentation, object recognition and image understanding
1/3/2014
instead of designing an algorithm, one could construct an example data set and an error criterion; the network input can consist of pixels or measurements in images; the output can contain pixels, decisions, labels, etc.; ANNs can be highly nonlinear; the amount of nonlinearity can be influenced by design, but also depends on the training data; some types of ANN have been shown to perform universal classification or regression techniques;
4
Feed Forward ANNs consist of interconnected layers of processing units or neurons. Basically, a feedforward ANN is a (highly) parameterized, adaptable vector function, which may be trained to perform classification or regression tasks. A classification feed-forward ANN performs the mapping N: Rd ( rmin, rmax )m, Self Organising Maps: Training is called competitive: at each time step, one winning node gets updated, along with some nodes in neighbourhood. After training, the input space is subdivided into q regions, corresponding to the q nodes in the map.
1/3/2014
An Adaline network is trained to perform an electrical impedance tomography (EIT) reconstruction, i.e., a reconstruction of a 2D image based on 1D measurement on the circumference of the image. Hopfield consists of a number of fully interconnected binary nodes, which at each given time represent a certain state. Connected to a state is an energy level The HNN maps binary input sets on binary output sets; it changes its state until the energy level is minimized. A Cellular Neural Networks is a system in which nodes are locally connected . Each node contains a feedback template and a control template, which to a large extent determine the functionality of the network.
5
the generalized adaptive neural filter (GANF) which has been used for noise suppression
any operation of which the input consists of sensor data, and of which the output is a full image.
Preprocessing operations generally fall into one of three categories: o image reconstruction (to reconstruct an image from a number of sensor measurements) o image restoration (to remove any aberrations introduced by the sensor, including noise) o image enhancement (accentuation of certain desired features, which may facilitate later processing steps such as segmentation or object recognition
IMAGE RECONSTRUCTION:
complex computations and a unique approach a reconstruction of a 2D image based on 1D measurements on the circumference of the image using ADALINE trained a modified Hopfield network. Having managed to map the problem appropriately, the HNN can be a useful tool in image pre-processing
IMAGE RESTORATION: employ all information about the nature of the distortions introduced
The restoration problem has a conflicting criteria which needs to be fulfilled: resolution versus smoothness regression feed-forward network in a convolution-like way to suppress noise while some built a modular feed-forward ANN approach that mimics the behavior of the filter, an edge-preserving smoothing filter.
1/3/2014
IMAGE ENHANCEMENT:
amplify specific (perceptual) features well-known enhancement problem is edge detection used a novel feed-forward architecture to classify an input window as containing an edge or not. The weights of the network were set manually to train a Hopfield network for enhancement of borders a general approach for implementation of morphological image operations by a modified feed-forward ANN using shunting mechanisms and a center-surround 7 shunting feed-forward ANN for contrast enhancement and colour night vision.
most important applications of data reduction are image compression and feature extraction
an image compression algorithm, used for storing and transmitting images, contains two steps: encoding and decoding
kind of features one wants to extract often correspond to particular geometric or perceptual characteristics in an image (edges, corners and junctions), or application dependent ones, e.g., facial features
8
1/3/2014
IMAGE COMPRESSION:
different types of image compression approaches can be identified: direct pixel-based encoding = decoding by one ANN and pixel-based encoding = decoding based on a modular approach.
ANNs trained to perform image compression: feed-forward networks, SOMs, adaptive fuzzy leader clustering (a fuzzy ART-like approach), a learning vector quantifier and a radial basis function network
approaches either combine different ANNs to obtain the best possible image compression rate or they combine more traditional statistical methods with one or more ANNs
1/3/2014
FEATURE EXTRACTION:
special kind of data reduction of which the goal is to find a subset of informative variables based on image detail and is also a means for controlling the so-called curse of dimensionality
Neural Nets in Feature Extraction is used for: Subsequent automatic target recognition in remote sensing (accounting for orientation) and character recognition, Subsequent segmentation of food images and of magnetic resonance (MR) images, finding the orientation of objects (coping with rotation), Finding control points of deformable models, clustering low-level features by filters, Subsequent stereo matching, clustering the local content of an image before it is encoded For a supervised auto-associator ANN, the in-formation loss implied by the data reduction can be measured directly on the predicted output variables, which is not the case for 9 unsupervised feature extraction by the SOM
SEGMENTATION
partitioning of an image into parts that are coherent according to some criterion
1/3/2014
Hierarchical models are built of different ANN types, e.g. using a SOM to map the image data to a smaller number of dimensions and then using a feed-forward ANN to 10 classify the pixel
TEXTURE SEGRREGATION:
A segmentation task that is most frequently performed by feature-based ANNs is texture segregation, which is typically based on co-occurrence matrices wavelet features Multi-resolution features extracted from the wavelets spatial derivatives computed in the linear scale-space
1/3/2014
COLOR RECOGNITION:
The discrete quaternion correlation (DQCR) for the application of color pattern recognition If we use DQCR for color pattern recognition, then we can detect the objects that have the same shape, size, color and brightness with the reference pattern
CLUSTERING:
The K-means algorithm is an iterative technique that is used to partition an image into K clusters. This algorithm is guaranteed to converge, but it may not return the optimal solution Lloyd's algorithm is a heuristic for solving the k-means problem, as with certain combinations of starting points and centroids, Lloyd's algorithm can in fact converge to the wrong answer The quality of the final solution depends largely on the initial set of clusters, and11 may, in practice, be much poorer than the global optimum
OBJECT RECOGNITION
Object recognition is potentially the most fruitful application area of pixel-based ANNs, as using an ANN approach makes it possible to roll several of the preceding stages (feature extraction, segmentation) into one and train it as a single system.
1/3/2014
Object recognition consists of locating the positions and possibly orientations and scales of instances of objects in an image.
The purpose may also be to assign a class label to a detected object. ANNs have been trained to locate individual objects based directly on pixel data.
To map the contents of a window onto a feature space that is provided as inputs
12
TEMPLATE
MATCHING
Several types of ANNs can be distinguished: feed-forward-like ANNs, variants using weight sharing recurrent networks, the ART networks introduced by Grossberg mixtures-of-experts , (evolutionary) fuzzy ANNs, bi-directional auto-associative memories, hardware ANNs have been built for object recognition, finally, SOMs are occasionally used for feature extraction from pixel data, the output of the map is then propagated to a classifier. An interesting approach that performs object recognition is the neurally inspired what-and-where filter. when Template Matching is performed by teaching a classifier to recognise the whole object from pattern of pixel intensities, the complexity of the classifier grows exponentially. A special type of ANN that incorporates the scale in-formation directly in a 13 pyramidal form is the so-called higher-order ANN.
1/3/2014
FEATURE
BASED RECOGNITION
variations in rotation and scale are coped with by the features, e.g., statistical moments
1/3/2014
certain amount of noise will influence the computed features and deteriorate the recognition performance
When the object to be detected is large and needs to be sampled densely, feature extraction is inevitable. Otherwise, a neural classifier will contain so many parameters that a good generalisation will be impeded. For object recognition, the features typically capture different local geometric properties. In situations where many degrees of freedom occur (2D or 3D rotation, scale, a One grey-level transformations, changes in colour, etc.), only feature-based approaches are used. When there is no limited set of images on which an algorithm has to work (e.g., image database retrieval), the more flexible pixel-based methods can prove useful 14
IMAGE UNDERSTANDING
couples techniques from segmentation or object recognition with the use of prior knowledge of the expected image content
ANNs were used in combination with background knowledge to classify objects such as chromosomes from extracted structures and to classify data which were recognised from pixel data. ANNs were used to analyse camera images for robot control from local features.
Neural (decision) trees, semantic models based on extracted structures or neural belief networks can be used to represent knowledge about the expected image content.
when applying ANNs for high level image understanding it becomes impossible to explain why a particular image interpretation is the most likely one. As remedy the trained ANN is mapped onto a Bayesian belief network or use the generic explanation facility developed for ANNs or to use 15 rule extraction
OPTIMISATION
1/3/2014
Some image processing (sub)tasks such as stereo matching can best be formulated as optimisation problems, which may be solved by HNNs.
Hopfield ANNs have been applied to the following optimization problems: SEGMENTATION OF AN IMAGE WITH AN INTENSITY GRADIENT BY
CONNECTING EDGE PIXELS
THRESHOLDING IMAGES BY RELAXATION TWO-DIMENSIONAL AND THREE-DIMENSIONAL OBJECT RECOGNITION BY (PARTIAL) GRAPH MATCHING ESTABLISHING CORRESPONDENCE BETWEEN STEREO IMAGES
BASED ON FEATURES AND STEREO CORRESPONDENCE BETWEEN LINE CAMERAS FROM DETECTED EDGES
1/3/2014
THANK YOU
17