Minoru Inamura

Improvement of remotely sensed low spatial resolution images by back-propagated neural networks using data fusion techniques

In this work an application of the back-propagated neural networks in the spatial resolution improvement of remotely sensed low resolution images using data fusion principles is described. Various tests are performed. In one study, the resolution of visible and near-infrared band images is enhanced using a panchromatic high resolution image and in the other, the resolution of the thermal band image is improved using other spectral band high resolution images. Different improvement rates are tested and the improvement covers spectral bands with wavelengths ranging from 0.45 μm to 2.35 μm and also the thermal band. The tests developed are explained and examples of the results obtained from each test are shown and compared with the correct patterns from which an error analysis is also carried out.

Spatial resolution improvement of remotely sensed images by a fully interconnected neural network approach

In previous works, backpropagation neural networks (BPNN) had been applied successfully in the spatial resolution improvement of remotely sensed, low-resolution images using data fusion techniques. However, the time required in the learning stage is long. In the present paper, a fully interconnected neural network (NN) model, valid from the mathematical and neurobiological points of view, is developed. With this model, the global minimum error is reached considerably faster than with any other method without regarding the initial settings of the network parameters.

Spatial resolution improvement of remote sensing images by fusion of subpixel-shifted multi-observation images

Multi-observation of satellite remote sensing provides the ability to achieve a higher spatial resolution image. Based on the relation between sensors of different spatial resolutions, this paper presents a multi-observation in spatial, called subpixel-shifted multi-observation, to acquire a more accurate image of higher spatial resolution than the original observations. In this kind of observation, the same area on the ground is observed repeatedly with a spatial resolution in a subpixel shifted way. All the acquired observation images are combined into a higher resolution image. This is formulated as a super-resolution equation. When comparing the existing super-resolution algorithms, we find that the Iterative Back-Projection (IBP) method suggested by Peleg et al. is an appropriate and effective method for solving this problem. Based on IBP, a pratical implementation is presented. Computer experiments on remote sensing images and error analysis show its effectiveness. Some problems, such as back-projection, undersampling, and fusion of observed samples, are discussed further. The resultant image from this method has both better quality and higher spatial resolution than the original observation.

Exterior Algebraic Processing for Remotely Sensed Multispectral and Multitemporal Images

This paper describes remotely sensed multispectral and multitemporal image processing from an algebraic point of view. Especially, image analysis by means of an inner product, an exterior product, and an inner product between two exterior products are presented.

Super-resolution of the undersampled and subpixel shifted image sequence by a neural network

Numerous approaches to super‐resolution (SR) of sequentially observed images (image sequence) of low resolution (LR) have been presented in the past two decades. However, neural network methods are almost ignored for solving SR problems. This is because the SR problem traditionally has been regarded as the optimization of an ill‐posed large set of linear equations. A designed neural network based on this has a large number of neurons, thereby requiring a long learning time. Also, the deduced cost function is overly complex. These defects limit applications of a neural network to an SR problem. We think that the underlying meaning of the SR problem should refer to super‐resolving an imaging system by image sequence observation, instead of merely improving the image sequence itself. SR can be regarded as a pattern mapping from LR to SR images. The parameters of the pattern mapping can be learned from the imaging process of the image sequence. This article presents a neural network for SR based on learning from the imaging process of the image sequence. In order to speed up the convergence, we employ vector mapping to train the neural network. A mapping vector is composed of some neighbor subpixels. Such a well‐trained neural network has powerful generalization ability so that it can be used directly to estimate the SR image of the other image sequences without learning again. Our simulations show the effectiveness of the proposed neural network.

Multidimensional filtering approaches for pre-processing thermal images

Demand for sharpened thermal images drives research into pre-processing techniques. This paper describes two fast multi-frame image-processing techniques for reducing noise and some blurring effects that are typically exhibited in thermal images. The first technique cleans the thermal image from random and fixed-pattern noises. The random noise is considerably reduced by the simple principle of averaging corresponding pixels of a multi-frame sequence. For eliminating fixed-noise like effects, the technique performs, at first, conventional arithmetic mean filters within each local region of the noise pattern. Then, weighted versions of these values are subtracted from the corrupted image. The second technique attempts to recover the information hidden at a sub-pixel level. It sharpens the previously processed thermal image by down-sampling and matching a set of sub-pixel shifted frames, and finally calculating the statistical weighted average within the correspondent aligned pixels of the multi-frame set. Some variants that combine it with conventional filters are also presented. This technique effectively corrects some blurring effects typically found in thermal infrared images. For the case of a single frame image determines the direction and width of the blur slope and re-assigns the max and min values to the correspondent pixels in the gradient direction. Then, the area is shifted and the same process is done again, up to cover the full image. Image evaluation methods demonstrate the accuracy and quality of the results. In addition to reducing the hardware requirements of present designs, these algorithms increase the utility of present sensors.

Spatial resolution improvement of a low spatial resolution image using spatial component extracted from high spatial resolution images

This method consists of two parts: extraction of the high spatial resolution from high spatial resolution visible and near-infrared images, reconstruction of the high resolution image from a low resolution image using extracted high resolution spatial information. Category decomposition for the high spatial resolution images extracts only high spatial resolution information from high spatial visible and near-infrared images and the reconstruction of high spatial resolution image is reconstructed by category composition processing. Where category composition and category decomposition are an inverse operation to each other. Examples of the spatial resolution improvement of low spatial thermal image and MOS-1 MSR image are demonstrated.<<ETX>>

Filtered multiple observation image superposition

When an observation satellite scans the ground object, the images of the object are taken at different times from different angles because of swing of the observation orbit. Based on the principle of computed tomography and considering the properties of multiple observation, this paper presents a new approach, filtered multiple observation image superposition, to improve the spatial resolution of remotely sensed imagery. In this method, first, all of the lower spatial resolution images are expanded into the same size images as some expected higher spatial resolution image by interpolation. Second, every expanded image is enhanced by iterative unsharp masking to increase the high frequency components. Finally, all of enhanced images are superimposed into one higher resolution image by weighted average. The validity of this method is examined by error analysis and comparison.

Improvement of IKONOS images with estimation of spatial information by neural networks

Simple enlargement of high resolution images induces conspicuous ringing. This paper deals with the new improvement technique using neural networks (NN). This method uses a fractal feature of images, and its fractal dimension characterizes the limits of its abilities.

Spatial resolution improvement of spatial shift multi-observation images by neural network

In this paper, improvements on spatial resolution of the spatial shift multi-observation images are discussed. And a block of pixels-based artificial neural network is proposed for this purpose. This system makes full use of the spatial information to implement the superposition of multiple images. Its convergence and learning problems are also discussed. The effectiveness and the high performance of the proposed neural network are demonstrated by computer experiments, error calculation and comparison with other methods.