An Image Enhancement Approach to Achieve High Speed using Adaptive Modified Bilateral Filter for Satellite Images using FPGA

,


Introduction 1.Background
Satellite images are widely used for various real time applications. Mostly these images are used for the weather forecasting, weather monitoring and atmosphere condition. The other applications includes in metrology, agriculture, forestry, super resolution. These images are captured from the low resolution cameras which causes lower illumination problem in images. Surveillance cameras would not have the ability to take accurate images.It is because of the relative motion between camera and vehicle. In today"s scenario, resolution of image is critical issue and other issue which affects the intensity value including brightness, contrast of image and color variations. For resolution enhancement various approaches have been proposed by the researchers for enhancing the resolution. Most effective technique is interpolation based approach and which has grown attraction of the researchers in image processing techniques. In recent years, several variations have been introduced as technology is growing faster which capture images at very high resolution.

Problem Definition
During the acquisition of satellite images, these images are corrupted by the noise which affects the quality of image. This degraded quality image causes analytical issues. This is the main challenge for the researchers to perform on-board operation in an efficient way to remove the noise or enhance the quality of image.
Satellite images are affected due to various aspects such as wrong setting of ISO, vibrations, motion in cloud and heat generation etc [2]. There are various other methods used Keeping the fine features of noisy image during the denoising process is crucial for the researchers to achieve the better results. Bilateral Filter is one of the effective denoising algorithms for removing unwanted noise while preserving the edges and this algorithm receives two parameters from the user. The user must select the most suitable parameter values to achieve the most meaningful result. Digital images procured through numerous gadgets items are regularly debased by Impulse commotion [2]. Images are regularly defiled by impulse noise because of errors produced in noisy sensors or channels. It is imperative to dispose of noise in the images before some processing, for example, detection of edges, segmentation and detection of objects. For this reason, numerous methodologies have been proposed [3]. It is understood that noise in the pictures may be presented amid the procedure of collection or transmission. Accordingly, image de-noising i.e. expulsion of noise from a picture is a critical piece of image processing [4]. Image processing is a procedure to improve images that is Edge/points of interest protection [7]. Image processing algorithms need to handle extensive measure of information. Programming usage will be tedious. For a few frameworks requiring ongoing handling, execution rate is frequently considered as a key component, so the image processing is suitable to be actualized in hardware. Field programmable gate array (FPGA) is suitable for pipelining and parallel information handling [1]. Prasetio, Barlian Henryranuet al [23] presented an encryption approach using FPGA to show the advantages of the FPGA kit.

Contribution of the paper
In this work we propose a new design for satellite image filtering or image enhancement to improve the quality of image by removing noise or unwanted signals [4].
During the image filtering, information related to images is lost due to the obscureness of sharp edges of image. In this work our main aim is to preserve the edges of image which provides better reconstruction. For image denoising median filter is a promising technique which is used widely for filtering operations. Median filter is an efficient technique which can remove the impulsive noise and salt & pepper noise [1]. Implementation complexity is also very less for median filters.
Still there are various challenges when median filters are implemented for satellite image system. To overcome these issues for satellite image denoising, here in this work we propose a modified bilateral filtering approach. This approach uses kernel based approach combined with bilateral filtering technique.
The main contribution of this approach is to perform the denoising operation at high speed with lower resource utilization and efficient reconstruction. In the results section we demonstrate the performance of the proposed approach in terms of operating frequency, device utilization summary and image reconstruction parameters.

Related Work
This section describes about the relevant studies which are carried out by other researchers in the field of image denoising. According to the study presented in [9] median filtering is a promising technique in the field of noise removal applications for signals and image processing schemes. Experimental studies presented shows that median filtering scheme is an optimal solution which is a significant approach for noise removal and information preserving. Median is computed by considering the middle element of group of a pixel. During pixel sorting in median filtering, all odd numbers of pixel elements are considered for computation. Figure 1 explains about the working steps of this approach by considering pixel window.
In [10] Zhu et al. proposed another approach for median filtering using cross window for median pixel computation. According to Fu et al [11] sorting is a key component of median filtering approach. To address this component, in [11] sorting is applied for pixel columns, diagonal elements and rows in a 3x3 window. If the size of window is considered as 3 then, operations which compares the columns requires 9 iteration and rows also needs 9 iterations for comparison.
In order to improve the performance of image filtering, adaptive filter is proposed which is capable to encounter the issues related to impulsive noise. Moreover, the adaptive median Pingjun Wei et al. [12] presented another study based on the adaptive median filtering. In their study they show that median filtering can perform image filtering if rectangular blocks are provided as an input. Novelty of other filters is that, it can adapt the window size during filtering operation which can vary based on the conditions.
An alternate methodology for the FPGA usage of a realtime two-sided channel has been proposed in [13]. The altered channel depends on the figuring of the channel coefficients from the photometric channel as it were. The spatial filtering is dispensed with because of the handling of the negligible window of 3 × 3 and raising of the determined photometric coefficients to the force of 8. As indicated by the researchers, for a moderate noise level, their altered reciprocal channel can accomplish marginally better results contrasted with the customary respective channel appeared in [14]. In any case, the first two-sided channel can be tuned by two parameters which are exceedingly in charge of the separating execution. Shockingly, no portrayal of the parameters utilized for this examination is given in [13].
Learning based approaches also being used by various researchers. Neural network is grown its significant in this field. In [15] Rezvanian et al. introduced two-pass scheme to remove impulse nose.
This scheme uses two approaches i.e. switching based scheme and progressive method. According to switching scheme, impulse detection approach is applied before filtering the image. This helps to obtain only noisy data for filtering. According to progressive approach, impulse noise detection and filtering are applied progressively using iterative approach. Experimental study presented in this work proves the effectiveness of approach for more corrupted images.
Al-Araji et al. [16] has implemented a new scheme for noise reduction by using occurrence rate estimation. Median filters and linear filters are main approaches for filtering the image. Median filters are applied for filtering so that the edge information can be preserved when compared to the linear filters [7].
Mainly impulsive noise can be classified into two main types: (a) salt and pepper noise and (2) random noise. According to salt & pepper noise either minimum pixel value or maximum pixel values are considered for processing whereas random valued noise can take any value of pixel which is complex when compared to salt and pepper noise.
Customarily, the salt-and-pepper noise is evacuated by applying median filtering scheme. At the point when the noise power is not exactly approx. 10% a straight forward middle using 3×3 or 5×5-pixel window is adequate. In any case, it was demonstrated that developmental configuration methods can create marginally preferable arrangements over standard medians for this noise power [2]. The images separated by developed channels are not all that smeared and the range on the chip can by diminished by approx. 60%. Among others, adaptive median filters give great results and at the same time with lower complexity in hardware and power consumption.

Background of Bilateral Filtering
This section describes about the state-of-art bilateral filtering scheme. This approach was first introduced by Tomasi and Manduchi in 1998 [3]. According to bilateral filtering consideration it is assumed that close pixels in image have same as spatial domain which concludes that these pixels will have similar intensity value. Filtering range is defined based on the weight decay with the dissimilarity values. Weights of image depend on the intensity of image.
As discussed, range filters are the type of non-linear filters.
In bilateral filtering scheme, both the filters domain filters and range filters are combined which results in better smoothing and edge preserving of image.
Due to its nature of edge persevering nature of bilateral filter, it is used in various applications for image denoising. Such as Adobe Photoshop use bilateral scheme for surface blurring tool. Similarly, GIMP implements selective Gaussian blur using bilateral filtering.
Another various application also uses bilateral filtering methodology such as digital still camera, surveillance camera and smart phones etc.
To improve the performance of the bilateral filter, we use kernel based approach for image smoothing. A Gaussian kernel is easy to implement which can improve the filtering performance and optimizing the speed tradeoff. Use of kernel is advantageous because it can perform averaging operation even when the pixel weights allocation is small.
Selection of kernels is carried out according to the requirements of image smoothing and computational requirements.
Conventional bilateral smoothing computes weight pixels by considering spatial distance from the center pixel.
By using this constraint, domain filter weight pixels are computed based on the distance follows: (1) and denotes the spatial position of the pixels and scaling of these pixels is fixed by using . Hence, by considering range filter weights, photometric difference can be computed as is the intensity of the image and filtering is fixed by the . By using equation 1 and 2, bilateral filter can be defined as In [17] Weijer et al. proposed a generalized approached for bilateral approach using space-tonal convolution method.
This technique addresses the basic problem of image smoothing and edge preserving of bilateral filters. If noise is added to any image, then the centered pixel also get affected which is used as a reference pixel for image filtering using tonal approach.
Work presented in [17] used the process of replacing the center reference value by estimating other true values of pixel. In [18]  detection of noise reduction of noise using anisotropic [18]. In their approach they suggested a combined approach for image denoising using low pass filtering with kernel function at a smaller extent.

Modified Bilateral Filter
In this section we discuss about modified bilateral filter approach for satellite image denoising using kernel functions. Conventional methods compute the weighted average from the neighborhood pixel but still some noise content remains present there. So there is a need to implement a new approach for removing noise after applying bilateral filter. To address this, we implement these modifications which are as follows: 1. Employing the replacement of summation. 2. Defining new kernels based on the knowledge of domain and range filters.
These modifications provide various advantages during implementation of denoising. In this work, for satellite image denoising application, we use Gaussian kernel based approach for computing the optimal weights. The computation process of Gaussian kernels is denoted in equation (4). To optimize the bilateral filtering the kernels are used which are given below:

Noise Removal using Modified Bilateral Filter
In previous sections we have discussed about the proposed approach of image filtering and noise removal. This section deals with detection of noise and removal using proposed approach.
Let us consider that an unknown image signal is denoted as a vector and degraded due to the Gaussian noise mixture into this. This is called a noisy image which can be represented as (8) In this work, our main aim is to remove the noise from input signal and restore it as an original signal. According to bilateral approach weighted average of pixels are compute which is given as This equation provides a normalized average value from the neighboring pixels from [ ] neighborhood. Computation of weights is carried out based on the content of neighboring pixels.

Image Quality Assesment
In this section we describe the formulas used to compute the quality of reconstructed image. In our model we are using 5 parameters of quality matrices for the assessing the image quality. These parameters are: (1) PSNR (Peak signal to noise ratio (2) MSE (Mean squared error) (3) Average difference (4) Maximum difference (5) Average Difference. These parameters can be computed by using below given equations.

Results
The proposed architecture of the modified bilateral filter was implemented in VHDL and simulated with Xilinx system generator. A test image was filtered by Xilinx system generator simulation, and the filtered images were compared. The target device was a Xilinx Virtex 4 XC4VSX55 FPGA. For the satellite image higher embedded memories are require, to use higher embedded memories we have selected family. For proposed model of image denoising, this platform is ideally efficient to perform the desired operation. This platform provides efficient number of logic elements, DSP blocks and reasonable memory requirements which makes implementation easier with reduced complexity.  The test image of port captured by satellite shown in Figure 3(a) is an grayscale image with a size of 64 × 64 pixels. In figure 3(b) Gaussian noise added image is show and in figure 3(c) speckle noise added image is shown. In Figure 4 port image with Gaussian noise and denoise image are shown Table 1.

Conclusion
In this work we address image denoising problem by applying image filtering scheme for satellite image systems. Main aim of the work is to reduce the complexity by limiting the required hardware components and optimizing the operating speed. This method is implemented on FPGA using pipeline method which uses filtering on row and column operators with pipeline filtered architecture.
We have presented a high speed implementation of the bilateral filter, targeted towards satellite image denoising.
Proposed image denoising scheme uses pipeline and parallel architecture for FPGA which makes it fully configurable to achieve the desired results.
For performance improvement we use a kernel based scheme combined with bilateral filters. This approach results in number of flip-flop reduction and resource utilization which make it feasible to implement.
This architecture uses fully scaled BRAM which helps to improve the operating frequency of 270 MHz In this work we show that the5x 5 kernel size uses resources efficiently which can be implemented for medium size of FPGAs. Kernel architecture reduce the use of external storage to improve the performance of logic utilization and achieve higher clock frequency.