AU2013100575A4 - Efficient and Effective Compression Algorithm for Infrared (IR) Image Based on Wavelet - Google Patents

Abstract

Abstract Infrared (IR) imaging techniques have been used extensively for military and civilian purposes, in particular in dark conditions. However, there is an important and popular process always to be involved, which is enormous data to be stored and transmitted during various applications. Many algorithms have been proposed to improve the performance of the compression scheme. In this patent we extended a traditional wavelet algorithm to an IR image compression with updated technology. It is compared our proposed algorithm, in terms of performance, with others related methods based on the standard parameters. It is well known that the technology of lifting based Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) wavelet transform is very efficient and effective in image compression. In our proposed algorithm, this technology will be coupled with Set Partition in Hierarchical Trees (SPIHT) coding algorithm and entropy coding techniques. One of our contributions in this patent is to demonstrate the choice of decomposition level, which is playing a very important role in achieving superior wavelet compression performances. An IR image quality is assessed objectively by the standard parameters, such as compression ratio, peak signal-to-noise ratio (PSNR), mean structural similarity index (MSSIM). It is also evaluated subjectively by using perceived image quality. It is to be supported by the simulation results that for our proposed algorithm are superior to others such as the compression ratio has been significantly improved by 88%, together with highest PSNR values and MSSIM which are close to 1. It is also found that the best decomposition level and required bit rate per pixel for infrared (IR) images in our algorithm. Keywords: IR, PSNR, MSSIM, SPIHT, Entropy coding. Figure 1: Proposed image compression block diagram Spliting Lifting IScaling Figure 2: The lifting-based Wavelet Transform Figure 3: Lifting implementation of the analysis side of the CDF 9/7 filter bank

Description

1 DESCRIPTION OF THE ART [1] Infrared (IR) light is electromagnetic radiation with a wavelength longer than that of visible light, measured from the normal edge of visible red light at 0.74 micrometers, and extending conventionally to 300 micrometers. These wavelengths corresponding to frequency range of approximately 1 to 400THz, and include most of the thermal radiation emitted to objects near room temperature. [2] Infrared imaging has been used extensively for military and civilian purpose, in particular in dark conditions. Military applications include target acquisition, surveillance, and night vision, homing and tracking. On-military uses include thermal efficiency analysis, remote temperature sensing, short ranged wireless communication, spectroscopy, and weather forecasting. As another example, infrared astronomy uses sensor-equipped telescopes to penetrate dusty regions of space, such as molecular clouds; detect objects such as planets, and to view highly red shifted objects from the early days of the universe. [3] Data compression stands for compressing data or files containing data so that they can be stored in much less memory space that they had been stored in their original form. Data transmission or storage without compression has some impractical reasons: (1) the data handle by different digital environments is increasing at a rate now a day for image processing application; (2) the storing of digital data without compression would be tragedy; (3) the transmission is major concern in modern world because of more than 7500 Tera Bytes of data is being downloading and/or uploading in the Internet. [4] In this patent, we have tried to find out the best decomposition level of compressed for IR images and also increased different bit rate for the best level of decomposition in terms of mean square error (MSE), peak signal-to-noise ratio (PSNR), mean structural similarity index (MSSIM) and compression ratio (CR). [5] We have used the technology called lifting based on CDF9/7 to compress the test images by using Set Partition in Hierarchical Trees (SPIHT) algorithm with entropy coding. It has been investigated to enhance the image quality with different level of decomposition. For the best performance in the image compression, we have investigated the most common entropy coding techniques, such as Run-length encoding and Huffman coding. These compression methods are compared over the basis of compression ratio which is defined as follows: Compression ratio is the define as Compression Percentage= AIBx 100 (1) Or alternatively, Compression Percentage = (A - B) / A x 100 (2) where A =Number of Bytes in the original data set, B = Number of Bytes in the compressed data set. The average number of bits required to represent the data value for the single pixel of an image is referred as bits/pixels. [6] The diagram of the trail process IR images compression and image decomposition can be seen in Figure. 1. In this proposed dissertation work IR image compress using Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) with lifting structure and Set 2 Partition in Hierarchical Trees (SPIHT) will be implemented. The dissertation work can be carried out in the following steps in our algorithm: Step. Read the IR image on the workspace of the MATLAB. Step2. Convert the given color image into gray level image. Step3. Perform CDF9/7 wavelet transform to the IR image: from the decomposition process the coefficients can be extracted. Step4. Apply Set Partition in Hierarchical Trees (SPIHT) encoding combined with Huffman encoding and Run length encoding reduced the redundancy in the coefficient data. Step.5 Set Partition in Hierarchical Trees (SPIHT) with Huffman encoded coefficients is saved the compressed bit streams instead of image. Step.6 Apply decoding procedure, from the compressed bit stream data, using Set Partition in Hierarchical Trees (SPIHT) combined with Huffman decoding and Run length decoding, as well as inverse CDF 9/7 wavelet transform to reconstruct the images. Step.7 Calculate compression ratio, MSE, PSNR and the overall image quality MSSIM. Step.8 Display the results reconstruction 1, reconstruction 2, reconstruction 3, i.e., level 1, 2, 3, 4.. .20 (as we considered) and comments on the quality of images with original image. Step.9 The above procedure is repeated for consider different bit rate per pixels with a fixed level of decomposition and display the results and compressed images. Stepl0. The same process is repeated for various IR images and compares its performance. [7] The lifting scheme based a wavelet transform can be implemented as shown in Figure 2, by which it is obviously clear that our scheme can reduce the computational complexity. It also can be seen that only the decomposition part of wavelet transform (WT) is depicted in Figure 2, this is because the reconstruction process is just the reverse version of the one described in Figure.2. The lifting-based WT consists of three sections, namely splitting, lifting, and scaling modules. The WT itself can be treated as prediction-error decomposition. It can be found that the scheme provides a complete spatial interpretation of WT. In Figure 1, let X denote the input signal and XL1 and XH1 be the decompose output signals where they are obtained through the following three modules ([A], [B], and [C]) of lifting base inverse DWT, which can be described as below: [A] Splitting-In this module, the original signal X is divided into two disjoint parts, i.e., samples X(2n+1) and X(2n) that denotes all odd-indexed and even-indexed and odd-indexed samples of X , respectively. [B] Lifting-Lifting consists of three basic steps: Split, Predict, and Updating. A brief description of these three steps, [(1), (2), and (3)], is given below. (1) Split -In this stage the input signal is divided in to two disjoint sets, the odd (X[2n+1]) and the even samples (X[2n]). This splitting is also called the Lazy Wavelet transform. (2) Predict-In this stage the even samples are used to predict the odd coefficients. This predicted value, P(X [2n]), is subtracted from the odd coefficients to give error in the prediction. d[n]=X[2n+1]+P(X[2n]) (3) Here, d[n]s are also called the detailed coefficients.

3 (3) Update-In this stage, the even coefficients are combined with d[n]s which are passed through an update function, U (.) to give Cn] = X(2n] + U(d[n]) (4) [C] Scaling-A normalization factor is applied to d(n) and s(n), respectively. In the even-indexed part S(n) is multiplied by a normalization factor Ke to produce the wavelet sub band X. Similarly in the odd-index part the error signal d(n) is multiplied by Ko to obtain the wavelet sub band XH1. [8] The lifting scheme based on a wavelet transform can reduce the computational complexity. The Lifting scheme of the wavelet transform Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) can go through of four steps: two prediction operators ('a' and 'b') and two update operators ('c' and 'd') as shown it Figure.3. [9] For lifting implementation, Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) pair can be factorized into a sequence of primal and dual lifting. The most efficient factorization of the polyphase matrix for the 9/7 filter may be following as follows: 1 a(1±+Z-)] 1 0 1 C(1±+Z-) 0[K 0 1 0 1 [b( 1+Z) lj 0 1 [d(1+Z) lj1 0 1/K) (5) where a, b, c, d and K are irrational values approximately equal to the following values that we also assumed to meet our case: az -1.58613432, b z -0.05298011854, cz 0.8829110762, dz 0.4435068522, Kz 1.149604398. [10] Instead of using equation (5), the following equation (6) describes the four "lifting" steps and the two "scaling" steps with same parameter as follows:

L

0 (n)= Y(2n), H 0 (n)= Y(2n +1) H'(n)=H 0 (n)+(ax [LO(n)+L 0 (n+1)]) L(n)= L 0 (n)+ (b x [H (n)+ H (n -1)])

H

2 (n)= H(n)+(cx [L(n)+L(n+1)]) (6)

L

2 (n) L (n)+(d x [H 2 (n)+ H 2 (n -1)])

H

2 (n) H(n)=H(n K L(n) = Kx L 2 (n) where H and L represents the high and low frequency component of input signal or image respectively. [11] The synthesis side of the CDF9/7 filter bank simply inverts the scaling, and reverses the sequence of the lifting and update steps. Figure.4 shows the synthesis side of the filter bank using lifting structure to reconstruct of the signal or image. [12] Set Partition in Hierarchical Trees (SPIHT) is the wavelet based image compression method. It provides the highest image quality, progressive image transmission, fully embedded coded file, simple quantization algorithm, fast coding/decoding, completely adaptive, lossless compression, and exact bit rate coding and error protection. SPIHT makes use of three lists - the List of Significant Pixels (LSP), List of Insignificant Pixels (LIP) and List of Insignificant Sets (LIS). These are coefficient location lists that contain their coordinates. After the initialization, the algorithm takes two stages for each level of 4 threshold - the sorting pass (in which lists are organized) and the refinement pass (which does the actual progressive coding transmission). The result is in the form of a bit stream. It is capable of recovering the image perfectly (every single bit of it) by coding all bits of the transform. However, the wavelet transform yields perfect reconstruction only if its numbers are stored as infinite precision numbers. [13] For image compression technique, it is well known that three of the most common entropy encoding techniques would be the Huffman coding, run length encoding (RLE), and arithmetic coding (AC). We shall concentrate on the Huffman and RLE methods for simplicity. Run-length encoding is a form of data compression in which it encodes a run of bytes to the following 2-byte form: {byte, length}, with length representing the number of runs of a single byte. On the other hand, Huffman coding techniques collects unique symbols from the source images and calculates its probability value for each symbol and sorts the symbols based on its probability value. Further, from the lowest probability value symbol to the highest probability value symbol, two symbols combined at a time to form a binary tree. Moreover, allocates zero to the left node and one to the right node starting from the root of the tree. To obtained Huffman code for a particular symbol, all zero and one collected from the root to that particular node in the same order. [14] Let xi and yi be the i t h pixel in the original image x and degraded image y, respectively. The MSE and PSNR between two images are given by MSE= - (x - y) 2 (7) N ,_1 PSNR =10logj j (8) SMSE where N is the total number of pixels in the image and R is the maximum fluctuation of the input image data value. For 8 bit/pixel gray-scale images, R=255. [15] The structural similarity index (SSIM), we shall use it to estimate the quality of medical images, specifically the ones compressed by CDF 9/7, based on the hypothesis that the human visual system (HVS) is highly adapted to extract structural information. The basic spatial domain SSIM algorithm compares the brightness, contrast and structure between each pair of vectors, where the structural similarity index (SSIM) between two signals x and y extracted from the original and degraded images, respectively, the luminance, contrast and structural similarity between them are evaluated as 2pxpy + C1 l(x,y) 2 2 p +C2 c(x,y) 2 x u+ 2 ax +( 2 +C j'X 2) s(x,y)= A Here, p, a,, and axy are the mean, standard deviation, and cross-correlation, respectively. C 1 = (K1L)2

C

2 = (K2L) 2 , and C 3 = 0.5 C 2 with K1= 0.01 and K2 = 0.03.

5 [16] The SSIM index define as the product of three components which gives by the equation: SSIM(x, y) = l(x,y) x c(x,y) (9) [17] For application, we require a single overall measurement of the whole image quality. The MSSIM (Mean Structural Similarity Index) values demonstrate greater reliability with the visual quality. It is given by the following formula: I1M MSSIM(X, Y) SSIM (x, y) (10) where X and Y are respectively the reference and degraded images, xi and yi are the contents of images at the ith local window. M is the total number of local windows in image. [18] We have applied algorithm to compress IR images. We have chosen an infrared image size 500x500 encoded on 8 bits per pixel, (Figure.10 (a)). This image is taken from the open database. The importance of our work lies in the possibility of reducing the rates for which the image quality remains acceptable. Estimates and judgments of the compressed image quality are given by the PSNR evaluation parameters and the MSSIM similarity Index. [19] From our investigating, we would highlight that the quality of compressed image depends on the number of decompositions level. The number of decompositions determines the resolution of the lowest level in wavelet domain. If we use larger number of decompositions, we will be more successful in resolving important DWT coefficients from less important coefficients. The HVS is less sensitive to removal of smaller details. [20] Fig. 9 shows comparison of reconstructed IR image for different levels, namely 1, 2, 3, 4, 5, 6, 7 and 8 decompositions. In this example, CDF9/7 combined with SPIHT and Huffman and Run length coding is used. It can be seen that image quality is better for a larger number of decompositions. On the other hand, a larger number of decompositions cause the loss of the coding algorithm efficiency. Therefore, adaptive decomposition is required to achieve balance between image quality and computational complexity. MSE, PSNR, MSSIM, and compression ratio (CR) tend to saturate for a larger number of decompositions [Fig. 6&7]. For each MSE, PSNR, MSSIM & CR, these image characteristics represents the optimal number of decompositions. From these image characteristics we found best decomposition level for IR image compression. We have seen best decomposition level is 5 as shown bold font in Figure 9(b). This decomposition has shown the roll of significant for the IR image. It has highest value of PSNR, MSSIM which is close to 1, which is a better compression ratio as compared to other decompositions level. [21] This five stage wavelet transform out is demonstrated at Figure.5 which gives the processing data unit for the next steps where SPIHT algorithm is used for data encoding. [22] To show the performance of the proposed IR image compression method, we made a comparison in terms of image quality increased by the MSE, PSNR, MSSIM & CR curves represented in Figures.6, 7 & 8. [23] The bit rates per pixel (BPP) were in the range of 0.125 to 2 respectively, and were chosen no uniformly such that the resulting distribution of subjective quality scores was approximately uniform over the entire image. The MSE, PSNR, MSSIM and CR measurement results are obtained. Obliviously, MSE perform as very poorly in this case that's why PSNR values were enhanced by the 6 increased the bits/pixel. The MSSIM values exhibit much better consistency with qualitative visual appearance. Those results are highlighted in Figure 8. [24] By comparing the different values of MSE, Peak Signal to Noise Ratio (PSNR) and Mean Structural Similarity Index (MSSIM) & Compression ratio (CR) as per increased the bit rate, we displayed as in Figures.10. The effectiveness of our proposed method in terms of compressed image quality which shown good results. We have seen our compression technique with decomposition level 5 found good result compressed image quality and also achieve higher compression ratio for as shown in Figure 10. It is good compression ratio 87.5157% for 0.125 bits/pixels but lowest PSNR value 38.2875dB but highest PSNR value 76.6828dB as shown the lowest compression ratio and MSSIM value is 1 which means the image quality is well for 2 bit/pixels. As expected bits/pixels was increased, the compression ratio enhanced from 13% to 88% approximately for bottom. The experiment shows that the higher data redundancy helps to achieve more compression. This experiment shows that CDF 9/7 couple with SPIHT, Huffman coding and Run length coding achieves more compression 87.5157 % for this proposed compression algorithm. [25] From the above results this technique performed well for gray IR image but poorly performed when compressing generic color IR images. This compression technique can save time in infrared (IR) image transmission and achieving process. So this simple and efficient compression technique can very useful in the field of infrared image processing and transmission. [26] In this patent, IR images quality have been enhanced evidenced by Figures 6-10. Under our novel algorithm it has shown the compression steps in our compression technique. We used the lifting based on Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) with the SPIHT coding to have the more effective compression. Two entropy codes such as Huffman and Run Length coding are also used after above technology to make further enhancement. In our proposed image compression algorithm it is noted that the compression is efficient and effective to both text and images. The proposed method provides better compression ratio approximately 88%, highest PSNR values and the performance of MSSIM measurements. It is achieving as overall correct recognition rate as 99.30 % values for IR images and this is more suitable for this category of images. Thus, we conclude that the results obtained are very satisfactory in terms of MSE, PSNR, MSSIM as well as compressed IR image quality. In perspective, we aim to apply this algorithm to compress all types of data not only to store but also for communication purpose with lower cost.

Claims (6)

1. A method for used Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7) with lifting structure and Set Partition in Hierarchical Trees (SPIHT), which using 10 steps: Step. Read the IR image on the workspace of the MATLAB. Step2. Convert the given color image into gray level image. Step3. Perform CDF9/7 wavelet transform to the IR image: from the decomposition process the coefficients can be extracted. Step4. Apply Set Partition in Hierarchical Trees (SPIHT) encoding combined with Huffman encoding and Run length encoding reduced the redundancy in the coefficient data. Step.5 Set Partition in Hierarchical Trees (SPIHT) with Huffman encoded coefficients is saved the compressed bit streams instead of image. Step.6 Apply decoding procedure, from the compressed bit stream data, using Set Partition in Hierarchical Trees (SPIHT) combined with Huffman decoding and Run length decoding, as well as inverse CDF 9/7 wavelet transform to reconstruct the images. Step.7 Calculate compression ratio, MSE, PSNR and the overall image quality MSSIM. Step.8 Display the results reconstruction 1, reconstruction 2, reconstruction 3, i.e., level 1, 2, 3, 4.. .20 (as we considered) and comments on the quality of images with original image. Step.9 The above procedure is repeated for consider different bit rate per pixels with a fixed level of decomposition and display the results and compressed images. SteplO. The same process is repeated for various IR images and compares its performance.

2. A method according to claim 1 the process is with three level's equation, L4(n), L 1 (n), and L 2 (n) (and H(n)s) to work out H(n)/K.

3. A method according to claim 1 where the targeted image is under compressing at the decomposition level is controlled by the pixel level based on DWT coefficients.

4. A method according to claims 1 and 3, where a decision is made for the lifting method based on Cohen-Daubechies-Feauveau wavelets with the low-pass filters of the length 9 and 7 (CDF 9/7), together with the SPIHT coding, to have the more effective compression. Then two entropy codes such as Huffman and Run Length coding are used to make further enhancement.

5. A method according to claim 4, the process order is such that the lifting wavelet, then filtering, then coding and comprising.

6. A method of recoding of mother wavelets and predictions in the processes are: d[n] =X[2n+1] + P(X[2n]) and update-In processing: C[n] = X[2n] + U(d[n]).