Fundamental Steps in Digital Image Processing

An Image is defined as a two dimensional function f(x, y). (x, y) is the spatial coordinate (or location) and f is the intensity at that point. If x, y and f all are finite and discrete, then the image is said to be a digital image. Digital image consists of finite and discrete image elements called pixels, each of them having a location and intensity value. In digital image processing, we process digital images using a digital computer.

Digital Image processing is not a field which is used for only high-end applications. There are various fundamental steps in digital image processing. We will discuss all the steps and processes that can be applied for different images.

Classification

We can categorise the steps in digital image processing as three types of computerised processing, namely low level, mid level and high level processing.

Low Level Processing

Low level processing involves basic operations such as image preprocessing, image enhancement, image restoration, image sharpening, etc. The main characteristic of low level processing is that both its inputs and outputs are images.

Mid Level Processing

Mid level processing involves tasks like image classification, object identification, image segmentation, etc. The main characteristic of low level processing is that its inputs are generally images whereas its outputs are attributes associated with image which are extracted from it.

High Level Processing

High level processing involves making sense of ensemble of recognised object and cognitive tasks associated with computer vision.

Fundamental Steps in Digital Image Processing

Image Acquisition

Image acquisition is the first step in digital image processing. In this step we get the image in digital form. This is done using sensing materials like sensor strips and sensor arrays and electromagnetic wave light source. The light source falls on an object and it gets reflected or transmitted which gets captured by the sensing material. The sensor gives the output image in voltage waveform in response to electric power being supplied to it. The example of a situation where reflected light is captured is a visible light source. Whereas, in X-ray light sources transmitted light rays are captured.

The image captured is analog image as the output is continuous. To digitise the image, we use sampling and quantization where discretize the image. Sampling is discretizing the image spatial coordinates whereas quantization is discretizing the image amplitude values.

Image Enhancement

Image enhancement is the manipulation of an image for its specific purpose and objectives. This is majorly used in photo beautify applications. These are performed using filters. The filters are used to minimise noise in an image. Each filter is used for a specific situation. Correlation operation is done between filters and input image matrix to obtain enhanced output image in . To simplify the process, we perform multiplication in the frequency domain which gives the same result. We transform the image from spatial domain to frequency domain using discrete fourier transform (DFT) multiply with filter and then go back to spatial domain using inverse discrete fourier transform (IDFT). Some filters used in frequency domain are butterworth filter and gaussian filter.

Majorly used filters are high pass filter and low pass filter. Low pass filter smoothens the images by averaging the pixel of neighbouring value thus minimising the random noise. It gives a blurring effect. It minimises the sharpening edges. High pass filter is used to sharpen the images using spatial differentiation. Examples of high pass filters are laplace filter and high boost filter. There are other non linear filters for different purposes. For example, a median filter is used to eliminate salt and pepper noise.

Image Restoration

Like image enhancement, image restoration is related to improving an image. But image enhancement is more of a subjective step where image restoration is more of an objective step. Restoration is applied to a degraded image trying to recover back the original model. Here firstly we try to estimate the degradation model and then find the restored image.

We can estimate the degradation by observation, experimentation and mathematical modelling. Observation is used when you do not know anything about the setup of the image taken or the environment. In experimentation, we find the point spread function of an impulse with a similar setup. In mathematical modelling, we even consider the environment at which the image was taken and it is the best out of all the other three methods.

To find the restored image, we generally use one of the three filters - inverse filter, minimum mean square (weiner) filter, constrained least squares filter. Inverse filtering is the simplest method but cannot be used in presence of noise. In the Wiener filter, mean square error is minimised. In constrained least error filtering, we have a constraint and it is the best method.

Colour Image Processing

Colour image processing is motivated by the fact that using colour it is easier to classify and the human eye can easily see thousands of colours than shades of black and white. Colour image processing is divided into types - pseudo colour or reduced colour processing and full colour processing. In pseudo colour processing, the grey scale is applied to one colour. It was used earlier. Now-a-days, full colour processing is used for full colour sensors such as digital cameras or colour scanners as the price of full colour sensor hardware is reduced significantly.

There are various colour models like RGB (Red Green Blue), CMY (Cyan Magenta Yellow), HSI (Hue Saturation Intensity). Different colour models are used for different purposes. RGB is understandable for computer monitors. Whereas CMY is understandable for a computer printer. So there is an internal hardware which converts RGB to CMY and vice versa. But humans cannot understand RGB or CMY, they understand HSI.

Wavelets

Wavelets represent an image in various degrees of resolution. It is one of the members of the class of linear transforms along with fourier, cosine, sine, Hartley, Slant, Haar, Walsh-Hadamard. Transforms are coefficients of linear expansion which decompose a function into a weighted sum of orthogonal or biorthogonal basis functions. All these transforms are reversible and interconvertible. All of them express the same information and energy. Hence all are equivalent. All the transforms vary in only the manner how the information is represented.

Compression

Compression deals with decreasing the storage required to the image information or the bandwidth required to transmit it. Compression technology has grown widely in this era. Many people are knowledgeable about it by common image extension JPEG (Joint Photographic Experts Group) which is a compression technology. This is done by removing redundancy and irrelevant data. In the encoding process of compression, the image goes through a series of stages - mapper, quantizer, symbol encoder. Mapper may be reversible or irreversible. Example of mapper is run length encoding. Quantizer reduces the accuracy and is an irreversible process. Symbol encoders assign small values to more frequent data and is a reversible process.

To get back the original image, we perform decompression going through the stage of symbol decoder and inverse mapper. Compression may be lossy or lossless. If after compression we get the exact same image, then it is lossless compression else it is lossy compression. Examples of lossless compression are huffman coding, bit plane coding, LZW (Lempel Ziv Welch) coding, (PCM) pulse code modulation. Examples of lossy compression are JPEG, PNG. Lossy compression is ideally used in the world as the change is not visible to the naked eye and saves way better storage or bandwidth than lossless compress.

Morphological Image Processing

In morphological image processing, we try to understand the structure of the image. We find the image components present in digital images. It is useful in representing and describing the images' shape and structure. We find the boundary, hole, connected components, convex hull, thinning, thickening, skeletons, etc. It is the fundamental step for the upcoming stages.

Segmentation

Segmentation is based on extraction information from images on the basis of two properties - similarity and discontinuity. For example, a sudden change in intensity value represents an edge. Detection of isolation points, line detection, edge detection are some of the tasks associated with segmentation. Segmentation can be done by various methods like thresholding, clustering, superpixels, graph cuts, region growing, region splitting and merging, morphological watersheds.

Feature Extraction

Feature extraction is the next step after segmentation. We extract features from images, regions and boundaries. Example of feature extraction is corner detection. These features should be independent and insensitive to variation of parameters such as scaling, rotation, translation, illumination. Boundary features can be described by boundary feature descriptors such as shape numbers and chain codes, fourier descriptors and statistical moments.

Image Pattern Classification

In image pattern classification, we assign labels to images on the basis of features extracted. For example, classify the image as a cat image. Classical methods for image pattern classification are minimum-distance, correlation and Bayes classifier. Modern methods for the same purpose use neural networks and deep learning models such as deep convolutional neural networks. This method is ideal for image processing techniques.

Applications

1. In medical diagnosis, Gamma ray imaging, X-ray imaging, ultrasound imaging, MRI imaging is used to know about the internal organs and bones of our body.
2. In satellite imaging and astronomy, infrared imaging is used.
3. In forensics, for biometrics such as thumbprints and retina scan, digital image processing is used.
4. We can find defects in manufactured packaged goods using microwave imaging.
5. We can find information about circuit boards and microprocessors.
6. Using image restoration, we can identify the car number plates of moving cars from CCTV for police investigations.
7. Beautify filters are used in social media platforms which use image enhancement.
8. We can classify and identify images using deep learning models.

Conclusion

A picture is worth a thousand words. And the world is filled with beautiful pictures. To manipulate these images according to our needs is all digital image processing is. And we live in the world using advanced digital imaging processing in diverse fields.

What is the commonly used language for image processing?

Python, C/C++ with openCV, Matlab, Java.

samuelnadar1211

Follow

Improve

Article Tags :

• Electronics Engineering
• Digital Image Processing