SOFM for Image Compression Based on Spatial Frequency Band-Pass Filter and Vector Quantization

SOFM for Image Compression Based on Spatial Frequency Band-Pass Filter and Vector Quantization

Shadi M. S. Hilles (Faculty of Computer and Information Technology, Al-Madinah International University, Malaysia) and Volodymyr P. Maidaniuk (Vinnytsia National Technical University, Ukraine)
DOI: 10.4018/978-1-7998-1290-6.ch016
OnDemand PDF Download:
No Current Special Offers


This chapter presents image compression based on SOFM and vector quantization (VQ). The purpose of this chapter is to show the significance of SOFM with bandpass filter in process of image compression to increase compression ratio and to enhance image compression effectiveness. Image compression by SOFM model is presented and consists of three stages: The first is band-pass filter. The result experiments used Lena.bmp, girl256.bmp, and show compression in block image 16x16 given best compression ratio with a small signal-noise ratio (SNR).
Chapter Preview


Image coding and decoding methods have gain many role internet media and multimedia technology for digital camera, digital TV and web applications where is huge amount of data transmission via internet is needed to increase compression ratio by remove redundant bits to reduce data storage and increase data transmission, however, there are several methods of image compression, there are enhancing new method day by day, especially of included artificial neural network methods, the image compression has two categories, first is lossless compression and second is lossy compression there is a trade-off between the compression and the quality (Raju, S., & Bhairannawar, S. S. 2017), the human conceptual of vision system is not sensitive with some of reducing bit redundancy and with changing in an image intensities, with high compression ratio is allow much loss bits and less image quality. The methods of Image compression which are required for web application does not need high image quality due to recognition of human vision conceptual, and beside that, in example medicine images and satellites are required high accuracy of image quality, included huge storage for images, The big challenges in image compression are how to save image quality after compression process and to get high compression ratio.

in this chapter is describe self-organization feature map SOFM from Kohonen scientific which is un-supervisor learning neural network, however, SOFM is one of the most common unsupervised neural network which has two layers, input and output with weight coefficients in each neuron connection in figure 1 illustrated system architecture of Kohonen map, SOFM model (Hilles, S. M. 2018) and (Huang, Y. H. 2016) and (Wang, J. P., & Cheng, M. S. 2012) and (Park, Y., & Suh, I. H. 2014) and (Hilles, S. M., & Maidanuk, v. P. 2014)

Figure 1.

Simplified performance model of a critical access module Artificial Neural Network of Self-Organization Feature Map SOFM



The study of Image compression methods and algorithms shows several contributions have been published in order to reduce image redundancy and increase compression ratio. As it has mentioned above there are two categories of image compression method lossy compression methods and lossless compression methods, the first is used for natural images photo gallery such as jpeg which is based on discrete cosine transform DCT, zigzag methods and entropy coding has used Huffman coding method, other example of lossy approach is jpeg 2000 which is based on discrete wavelet transform and adaptive arithmetic coding for entropy, the second category of image compression is lossless which is common used in clip art, comic arts and technical drawings(Jasmi, R. P., Perumal, B., & Rajasekaran, M. P. 2015, . Hilles, S. M., & Hossain, M. A. 2018), presented classification of data compression method such as RLE, arithmetic coding, LZW and Huffman coding as lossless compression methods, it shows and classified lossless methods in image compression model where using transformation such as DCT and DWT and then entropy coding in the last stage, (Han, S., Mao, H., & Dally, W. J. 2015). The deep compression of using neural network presented method for trained quantization, the idea is based on three steps pruning, quantization and Huffman coding, and the quantization is actually comes from weights of neural networks.

Moreover, to use hybrid combination of image compression techniques such as SOFM and vector quantization with Set Partitioning in Hierarchical Trees SPIHT coding, the model shows wavelet transform as first stage of process compression (Rawat, C. S. D., & Meher, S. 2014). SOFM presented and investigated (Hilles, S., & Maidanuk, V. P. 2014), the comparative characteristic of fractal method with SOFM, and by using 16x16 Kohonen map SOFM shows high compression ratio coefficient with small arithmetic operations, the signal noise ratio SNR given small error when compare with JPEG, however, the result shows better performance of compression by using 16x16 Kohonen Map.

In addition, using band-pass filter for analyzing oriented edges to allow many edges as possible in preferred orientation band to pass, (Park, et al., 2014) the result evaluated proposed model with classical filter-based methods.

Complete Chapter List

Search this Book: