Image and Video Processing

As part of the ENEE 631 : Image and Video Processing couse conducted by Dr. K. J. Ray Liu at UMD, I had to implement and submit six lab reports and a term project report on various topics. These reports and codes will be uploaded here.

Notes:

1. All the implementations are in Matlab.

2. Run the files starting with main. In the code folders, there are pdfs named lab* which contain the original definition of the problem.

LAB-1

Image Enhancement through Contrast Stretching and Histogram Equalization

Abstract

In this report, image enhancement is performed through the modification of luminance histogram using two different methods. In the first method namely, contrast stretching, piecewise linear transformation of the gray level of the original image is performed to stretch the one-sided histogram to the whole greay-level region. In the second method namely, histogram equalization, the histogram is equally spread in all the gray-level bins by employing transformation of the cumulative distribution of pixel values. Outcomes of the two different methods are compared in this literature and their respective advantages and disadvantages are discussed.

Keywords: Histogram Equalization, Contrast Stretching, Gray Level Transformation.

Full Report is here.

Matlab code is here.

LAB-2

Edge Extraction from Noisy Image

Abstract

In this report, a methodology for edge detection from a noise corrupted image is discussed. The given image is corrupted with two different types of noises. The task therefore requires image denoising prior to employing edge detection. Median filter and gaussian smoothing filters are employed for obtaining the noise suppressed image. Then edge detection is performed using gradient operators operators. Different parametric combinations of these methods are tested on the given noisy image and the results are compared.

Keywords: Salt and pepper noise, gaussian noise, median filtering, image smoothing, edge detection, non maximum suppression

Full Report is here.

Matlab code is here.

LAB-3

Image Restoration

Abstract

In this report, a methodology for image restoration from a blurred image is discussed. The given image blurred image is the result of passing the original image through a gaussian smoothing filter. The task is to removed the blur effect and produce a sharp picture comparable to the original image. In this regard, two of the most popular deblurring methods namely, the inverse filtering and the Wiener filtering approaches are explored. The mean square errors of the resulting images with the original one are calculated and the performances of the two methods are compared.

Keywords: Image deblurring, inverse filtering, gaussian smoothing, wiener filtering

Full Report is here.

Matlab code is here.

LAB-4

Image Compression Using Discrete Cosine Transform

Abstract

In this report, a methodology for image compression for both color and grayscale images is discussed. The technique is a simpler version of the well known JPEG compression. Here, cosine transform on image-blocks is performed following segmentation of an image into blocks. Then, the cosine transform of each block is quantized using suitable number of bits to achieve the compressed image. The task is a bit more complicated for colored images and requires transformation into the YCbCr domain from RGB. Compression performance of the method on the two images are compared in terms of the image compression ratios while the qualities of the output images are evaluated in terms of peak signal-to-noise ratio (PSNR). A target PSNR value of $35$dB is set and the best compression ratio is achieved while maintaining the targeted PSNR.

Keywords: Image compression, cosine transform, quantization, compression ratio (CR), peak signal-to-noise ratio (PSNR)

Full Report is here.

Matlab code is here.

LAB-5

Image Compression Using Embedded Zero tree Wavelet Algorithm

Abstract

In this report, a methodology for image compression for both color and grayscale images is discussed. The technique is a simpler version of the JPEG-2000 compression algorithm which is known as embedded zero tree wavelet algorithm (EZW). The method is based on bi-orthogonal wavelet decomposition of the image followed the formation of embedded zero tree and finally successive approximation quantization of the wavelet coefficients. Compression performance of the method on the two sample images are compared in terms of the image compression ratios while the qualities of the output images are evaluated in terms of peak signal-to-noise ratio (PSNR). A target PSNR value of $35$dB is set and the best compression ratio is achieved while maintaining the targeted PSNR. In this report, the compression performance of the EZW is compared with that of the DCT based compression technique, used for JPEG compression in the past, in terms of the performance measures.

Keywords: Image compression, wavelet transform, embedded zero tree wavelet algorithm (EZW), quantization, compression ratio, peak signal-to-noise ratio (PSNR).

Full Report is here.

Matlab code is here.

LAB-6

Motion Estimation and Compensation

Abstract

In this report, a methodology for motion estimation and compensation for a video sequence is discussed. The technique employed here is the simplest block matching approach that performs exhaustive search to detect movement between two consecutive frames of a video. The main idea is to obtain the motion vectors that contain the shifting information of the best matching block of the reference frame in the predicted frame and also calculating the prediction error. Since, these two information are sufficient to reproduce the a frame from the reference frame, the method can effectively reduce the memory requirement to a bare minimum. In this report, the technique of motion estimation and compensation is discussed with two different examples.

Keywords: Motion Estimation, Motion Compensation, Mean Absolute Distance (MAD).

Full Report is here.

Matlab code is here.

TERM PROJECT

Sparse Representation and Recognition of Faces

Abstract

In this report, the task of face recognition in the presence of illumination, expression and occlusion is addressed. Here, the face recognition problem is modeled as a problem of classifying among multiple linear regression models by representing the test face images as a sparse combination of training images from the same class. It is shown that for a sufficiently large feature vector, the sparse representation can produce very high recognition accuracy compared to some of the state-of-the-art face recognition techniques especially in case of occluded faces. Moreover, it is shown here that a block based approach of face recognition can produce even better results. Moreover, the given framework can handle errors due to occlusion and corruption uniformly by exploiting the fact that these errors are often sparse with respect to the standard (pixel) basis. Aside from modeling the sparse representation problem as a basis pursuit denoising problem, in this paper, two other method for obtaining sparse solution, namely, the orthogonal matching pursuit and the group lasso are employed. It is shown that the later two methods are fast and effective. In order to improve the accuracy further, block based methods are adopted and the improvement in overall performance is demonstrated. The optimization algorithms are combined with feature dimensionality reduction techniques to obtain improved accuracy and faster implementation of the system. Extensive experimentation is performed on two separate datasets to verify the validity of the models.

Keywords: Sparse Representation, Dictionary Learning, Basis Pursuit Denoising (BPDN), Orthogonal Matching Pursuit (OMP), LASSO, Group LASSO, principal component analysis (PCA), linear discriminant analysis (LDA).

Full Report is here

Author: Upal Mahbub, umahbub [at] umd [dot] edu