Even under ideal conditions side-scan sonar (SSS) images of targets can vary greatly depending on the target range and orientation, even if their geometries are identical. This complicates target classification algorithms since typically only a small samples of targets are available for training purposes. This under-representation of targets can cause missed classifications and a higher false alarm ratio in the presence of clutter. This problem is addressed by using a priori information about the targets as well as the imaging system embedded in a model for simulating target images. These simulated target images can be added to the training set for a more complete target representation. Another important aspect of this research includes the use of multiple channels extracted from the images using a multi-resolution wavelet decomposition. This multi-resolution analysis is used to first provide for an efficient detection strategy, by filtering the images over the lower resolution channels. Furthermore, providing target features at different scales improves the performance of the neural network classifier. The dependence of the classifier on local image enhancement provided by total variation minimization and Mumford-Shah segmentation is also studied.
Mine detection using model-trained multi-resolution neural networks and variational methods
1999
11 Seiten, 13 Quellen
Conference paper
English
OBJECT DETECTION USING MULTIPLE NEURAL NETWORKS TRAINED FOR DIFFERENT IMAGE FIELDS
European Patent Office | 2022
|Nuclear Reactor Diagnostic System Using Genetic Algorithm (GA)-Trained Neural Networks
Online Contents | 1995
|