A Survey of Image Classification and Techniques for Improving Classification Performance

Abstract

In image analysis classification of land used image is an important application. There are various algorithms used for classification of data some algorithms are rule based and some algorithms are learning based. We may get good classification but some pixels are always misclassified or unclassified. The major reason for misclassification is mixed pixel. The composition of the various objects in a single pixel makes identification of genuine class more difficult. Subpixel algorithms give the better idea about the respective class of such pixels. The subpixel mapping method is varies depending on the type of image. In Panchromatic or multispectral images the data set is very less as compared to the hyperspectral image. A hyperspectral image contains contiguous bands. Each band is very narrow with few nanometer bandwidths. More than a hundred such bands are available in the hyperspectral image. This huge data set is very difficult for the typical neural network to process. The feedforward neural network is not able to reach the local minima whereas the back propagation neural network needs a lot of time to converge to a minimum value. Radial basis function neural network has some advantages over other but it gives poor performance on hyperspectral imaging. The convolutional neural network is going to resolve the huge data problem. It has a 3-dimensional vector in which we can take multiple kernels to operate on interested data. This kernel gives us depth which is nothing but the more information of the same pixel. So here we can save a lot of information as compared to other neural networks. But in the convolutional neural network after the pooling layer, our data is in a 3D form which we need to convert again in 1D by flattening.

Keywords: Convolutional Neural Network (CNN), Spectral Unmixing, Hyperspectral, Subpixel