Geri Dön

Parmakizi imgesinin yönsel süzgeçler ile iyileştirilmesi ve dalgacık dönüşüm kodlama yöntemi ile sıkıştırılması

Başlık çevirisi mevcut değil.

  1. Tez No: 55837
  2. Yazar: H.NURİ YANANLI
  3. Danışmanlar: Y.DOÇ.DR. M. ERTUĞRUL ÇELEBİ
  4. Tez Türü: Yüksek Lisans
  5. Konular: Elektrik ve Elektronik Mühendisliği, Electrical and Electronics Engineering
  6. Anahtar Kelimeler: Belirtilmemiş.
  7. Yıl: 1996
  8. Dil: Türkçe
  9. Üniversite: İstanbul Teknik Üniversitesi
  10. Enstitü: Fen Bilimleri Enstitüsü
  11. Ana Bilim Dalı: Belirtilmemiş.
  12. Bilim Dalı: Belirtilmemiş.
  13. Sayfa Sayısı: 47

Özet

ÖZET Bu çalışmada, parmakizi imgesi elde edilirken oluşan bozuklukları gidermek için iki farklı iyileştirme algoritması ve parmakizi imgesinin veri tabanında daha az yer işgal etmesi için FBI tarafından kullanılan bir sıkıştırma algoritması kullanılmıştır. İyileştirmede kullanılan yöntemler: Yönsel Fourier süzgeçleme ve kum saati tipi süzgeçler ile altbant süzgeçleme. Bu iki yöntem aynı yapıya sahip olup farklılık yönsel süzgeçlerin elde ediliş yöntemlerindedir. Orijinal imgeden hesaplanan sağrı yönelimlerinin ağırlıklı değerlerine göre yönsel imgelerin ortalaması alınarak iyileştirilmiş imge oluşturulur. Birinci yöntem olan yönsel Fourier süzgeçleme tekniğinde, iki tek boyutlu süzgecin çarpmamdan oluşan bir süzgeç kümesi kullanılmıştır. Böylece sağrı kalınlıklarına göre belirli yöndeki sağrılar süzülerek yönsel süzgeçlenmiş imge kümeleri elde edilmiştir. İkinci yöntem olan kum saati tipi süzgeçler ile altbant süzgeçleme tekniğinde ise iki- banth kum saati tipi iki-boyutlu süzgeçler kullanarak bir ağaç yapısı içerisinde yönsel süzgeç kümeleri elde edilmiştir. Kullanılan süzgeçlerin ideal frekans cevaplan olmadığından komşu bantlardan karışmalar meydana gelmektedir. Bunu engellemek için yönsel süzgeçlenmiş imgeler frekans düzleminde bir eşikten geçirilir. Parmakizi imgesinin sıkıştırılmasında ise dalgacık(wavelet) dönüşümü tekniği kullanılmıştır. Sıkıştırma üç adımdan oluşuyor. Birinci adımda dalgacık dönüşümü kullanılarak imge bantlara aynştinimış, ikinci adımda ise oluşan bantlar adaptif kuantalama tekniği ile kuantalanmıştır. Son aşamada bu kuantalama seviyelerine huffman kodlama kullanarak bir kod cümlesi atanmışur. Sıkışülmış parmakizi imgesi ters işlemler uygulanarak yeniden parmakizi imgesi elde edilmiştir. v

Özet (Çeviri)

SUMMARY FINGERPRINT IMAGE ENHANCEMENT with DIRECTIONAL FILTERS and COMPRESSION with WAVELET TRANSFORM CODING In the automatic fingerprints recognition system, the quality of original print is the major problem. If the quality is not an acceptable standard, automatic fingerprint identification becomes extremely difficult. The reason for this is that normal methods of fingerprint recognition use the small unique features (known as minutiae) in the fingerprint pattern to identify the fingerprint. However, it is extremely difficult to extract these minutiae from the fingerprint image if the quality of the print is not perfect. Enhancement algorithms are used to eliminate undesired information (as false minutiae) and to emphasize desired information (as right minutiae). In this study two different approach are applied to enhance the fingerprint images. First approach is directional Fourier domain filtering [1] and second is subband filtering by using hour glass shaped filters. General structure of enhancement algorithms is shown in Figure 2. 1 and the structure is the same as each other. The difference is the way of obtaining directional filters. Filters are applied to image, yielding a set of directional filter banks which are called prefiltered images. The filtered image is then built up by selecting, for each pixel position, the pixel value from the prefiltered image whose direction of filtering corresponds most closely to the actual ridge orientation at that position. In order to perfom this selection operation, knowledge of actual ridge orientation is required, and this is obtained by estimation from the original image. The first approach is based upon nonstationary directional Fourier domain filtering. Using aim of Fourier domain filtering and prefiltered images are to convolve the fingerprint images with filters of full image size, since the two-dimensional fast Fourier transform algorithm can be used to calculate convolution efficiently. Fingerprint exhibit everywhere a well defined local ridge orientation and ridge spacing. This enhancement algorithm takes advantage of this regularity of spatial structure by filtering the image with a position-dependent directional Fourier domain filter whose passband is everywhere matched to the local ridge orientation and spacing. Figure 2.5 illustrates this enhancement process. Choosing a representative set of quantified directions defining a set of directional filters which, when applied to the original image, yield a set of directionally filtered linages. In this way, an anisotropic nonstationary bandpass filter which depends parametrically on LRS (Local VIridge spacing) and LRO (Local ridge orientation) are defined. The polar coordinates are used to express the filter as a seperable function: H^^H radial^" a^,e^ and in order to allow independent manipulation of its directional and radial frequency responses, H^Mai (p) depends upon LRS and Ha,,gie () upon LRO. Unlike the first approach in the second method spatial convolution of the image with filter masks are used. This approach is shown in Figure 2.4. It is performed within directional subbands in a tree structure by using 2-Dimension hour-glass shaped filters. Spectrum is divided into two subbands in each step and this process is repeated twice to perform four directional subbands. Altough fingerprint images divided into four bands show LRO in certain directions, there are some interferences within the band from other directions because the designed filters do not have ideal frequency responses. In order to filter the interference, a threshold is applied to the bands within the frequency spectrum. Finite impulse response (FIR) filter structure is used in designing the filters. Directional filters are designed by using the 2-Dimensional Bernstein approach.The desired frequency response H(wi,w2) is obtained by first defining a function H(wi,W2) between 0 [1/2(1- cosw,), 1/2(1 - cosw2)] and then the function is transformed into frequency domain. Rotation is used to perform different directional filter banks. At the end, the tree-structure partitioned subband wedges are rotated to their original position. These two approaches are defined as bandbass filtering which depends on LRO. Values of LRO at each point are required; these are estimated from the image data. This filter passes the ridge information while eliminating most of the noise, since the noise is not localised in the same way as the ridge information. The value of LRO at each pixel is required as parametric input to the filter. Since determining LRO reliably can be computationally demanding, it may not be feasible to estimate LRO directly for every pixel. LRO is determined at a square grid spaced 8 pixels apart, intermediate values are obtained by interpolation. This window is rotated to n different orientations, Oi = i.7t/n for i=0..n-l In the first approach n is selected 8 and for the second approach n=16 orientations are used.At each orientation a projection along the y-axis of the window is performed: vn/?(*) = -E *K*jO x = o...«-ı »y=0 where Wi(x,y) is the data inside the window at angle Bj. Total variation V; of each projection is calculated for each direction as: H-l ^=5;k(*+i)-Ji(x) x=0 Considering the probability that there may be more than one direction in a window, the total variation Vi of each filtered projection is put through a threshold and the enhanced image is built after taking the weighted values according to Vi of the directional subbands. An alternative, equally acceptable approach is to use a faster but perhaps less reliable algorithm to estimate orientation at every pixel position and to smooth the resulting orientation image. The results which are obtained from directional filtering and the value of LRO are brought together to present an algorithm which produces high quality enhancement of fingerprint images. Directional filtering images original images and enhanced images are shown in Figure 4.1 and 4.2. The software is developed in C programming language and Matlab. Programs are represented in Appendix- A and Appendix-B. The second part of this study is fingerprint image compression. The main objective of the fingerprint image compression is to represent an image with as few bits as possible while preserving the level of quality the original image. First and most important element is the image decomposition or transformation is performed to eliminate redundant information, and to provide a representation that can be coded more efficently. In this study discrete wavelet transform is used for decomposition. Lowpass and highpass linear-phase finite impulse response (FIR) filters are used. The filters contain pairs of odd-length, symmetric (impulse responses are symmetric about their middle sample). These are called Type I lineer phase FIR filters or whole-sample symmetric (WSS) filters. The impulse response coefficients for the analysis filters are given in Table 3.1. The synthesis filters are completely determined by their anti-aliasing relations. Before filtering stage the input image is symmetrized due to symmetric properties of filters. Symmetric filters are applied to the signal by computing circular convolution of the signal and the filter. Then the input signal is divided into a set of uncorrelated frequency bands by using wavelet linear phase FIR filters. At the end of analysis stage, the frequency spectrum is split into 64 subbands. The subband tree structure is shown in Figure 3.10 and the filter bank path for this tree structure is given in Table 3.3. VIIIThe second element of image coding is quantization. To represent an image with a finite number of bits, transform coefficients are quantized. After the subband decomposition is computed the resulting coefficients are quantized uniformly within subbands. The subbands are quantized proportionately to the log of the energy level of the subband with an additional adjustment to certain high frequency bands containing esseantial identification features. The quatization step size depends on quantization factor. Then the quantization factor controls the compression rate and quality of the reconstructed image. The smaller quantization factor means the higher compression rate and worse reconstructed image. The third element in the image coder is assignment of codewords, the string of bits that represent the quantization levels. Consecutively numbered subbands are combined into blocks for entropy coding. Then quantization levels are represented by symbols as shown in Table 3.4. This symbols are coded by using Huffman encoder. Reconstruction is achieved by decoded and dequantized subbands, applying appropirate filters, and adding the reconstructed subbands together. The software are developed in Matlab. Huffman coder is written in C programming language. Programs are represented in Appendix- C. In evaluating the reconstructed image quality provided each implementation, root- mean squared error (RMSE) and peak signal-to-(reconstruction) noise (PSNR) are used as error metrics. The other metrics for reconstructed image quality is error image. Error images, representing the difference between the original and reconstructed image. For different compression rate, the original image, reconstructed image and the difference between both which is called error image are shown in Figure 4.9, 4.10, 4.11, 4.12 and 4.13. This algorithm is applied in FBI and it is accepted as a standard for fingerpint image transmission. Results are given below according to quantization factor: LXTable: Compression rates for Wavelet Scaler Quantization technique according to quantization factor.

Benzer Tezler

  1. Tez No

    541786

    Deep convolutional neural network based unconstrained ear recognition

    Derin evrişimsel sinir ağı tabanlı kısıtsız kulak tanıma

    FEVZİYE İREM EYİOKUR

    Yüksek Lisans

    İngilizce

    İngilizce

    2018

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    DOÇ. DR. HAZIM KEMAL EKENEL

  2. Tez No

    503382

    Face recognition with local Walsh transform

    Yerel Walsh dönüşümü ile yüz tanıma

    MERYEM UZUN PER

    Doktora

    İngilizce

    İngilizce

    2018

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. MUHİTTİN GÖKMEN

  3. Tez No

    139584

    Biometric identity verification using on-line & off-line signature verification

    Dinamik ve statik imzaları kullanarak biyometrik kimlik doğrulama

    ALİSHER KHOLMATOV

    Yüksek Lisans

    İngilizce

    İngilizce

    2003

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolSabancı Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    YRD. DOÇ. DR. AYŞE BERRİN YANIKOĞLU

  4. Tez No

    741022

    Presentation attack detection with shuffled patch-wise binary supervision

    Karıştırılmış yama tabanlı ikili gözetim ile sunum saldırı tespiti

    ALPEREN KANTARCI

    Yüksek Lisans

    İngilizce

    İngilizce

    2022

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve Kontrolİstanbul Teknik Üniversitesi

    Bilgisayar Mühendisliği Ana Bilim Dalı

    PROF. DR. HAZIM KEMAL EKENEL

  5. Tez No

    539080

    Algısal özet fonksiyonları tabanlı derin öğrenme yöntemleri kullanılarak imgelerin sınıflandırılması

    Classification of images by using deep learning methods based on perceptual hash functions

    FATİH ÖZYURT

    Doktora

    Türkçe

    Türkçe

    2019

    Bilgisayar Mühendisliği Bilimleri-Bilgisayar ve KontrolFırat Üniversitesi

    Yazılım Mühendisliği Ana Bilim Dalı

    PROF. DR. ENGİN AVCI

Geri Dön