Image Processing with neural networks-a review
M.Egmont -Petersen,D.de Ridder,H.Handels, Pattern Recognition,Vol 35, Issue 10, Oct 2002, Pg: 2279–2301.
This paper reviews the various applications of neural networks developed to solve different problems in image processing. The problems addressed in the field of image processing is organised into a image processing chain. This chain includes the following:
- Preprocessing/Filtering - It enhances the contrast and helps in noise reduction. It includes 3 categories: Image reconstruction, image restoration and image enhancement. Artificial Neural Networks(ANN) were applied directly to pixel data. Eg: Cellular Neural Networks(CNN), Generalized Adaptive Neural Filter(GANF), hopfield network
- Data reduction/Feature extraction- It extracts components from the image. Features are smaller than the number of pixels. The application includes image compression for storage and transmission and feature extraction for object recognition and segmentation. ANN is used for both. Adavantage of using ANN is that the parameters are adaptable giving better compression rates when trained on specific image material. Eg: feed-forward networks, Self-Organising feature Map (SOM), adaptive fuzzy leader clustering, hopfield.
- Segmentation - It partitions an image into regions. Neural networks perform segmentation directly on the pixel data obtained from a convolution window or the information is provided to a neural classifier in the form of local features. Eg: feed-forward networks, SOM, hopfield networks, probabilistic ANNs.
- Object detection and recognition- classifies the object and determines the position, orientation and scale of specific objects in the image.ANN were trained to locate individual objects based directly on the pixel data and also based on features that is provided as input to the neural classifier.
- Image Understanding- couples techniques from segmentation and object recognition with knowledge of expected image content.ANN were used to classify objects, analyse images to extract knowledge about image content.
- Optimization- optimizes the function for graphing and object extraction. Eg: Hopfield network.