LetterNon-uniform image compression using a biologically motivated selective attention model
Introduction
In general, there exist two standard image compression methods [6]. One compresses an image by a lossless coding algorithm. A lossless compression method does not suffer from the missing problem of information, which means there is no decoding error. However, the compression ratio is not so high. The other method is lossy compression. Its compression ratio is very high, but a loss of information inevitably occurs. Even though the lossy compression method is efficient, important information may be lost because the uniform compression algorithms compress all contents equally. If we consider non-uniform image compression that compresses the contents in an image according to the relative importance in a different way, we are able to realize a more efficient image compression coding scheme.
Non-uniform image compression has been applied to many engineering domains such as mobile applications, biomedical applications, and tele-browsing [9], [11]. However, it is difficult to discriminate between the regions that will be compressed by a lossless coding algorithm and those that will be compressed by a lossy coding method. However, the human visual attention system can focus on an attentive location in an input scene and select interesting visual information to process in the brain. Additionally, humans can ignore an uninteresting area even if it has salient primitive features, and they can memorize the characteristics of the unwanted area. Therefore, we propose a new non-uniform image compression method using a human-like selective attention mechanism that considers not only primitive input features, but also the interaction with the environment. The selective attention model can indicate a meaningful area in a natural image. The most salient point is used as the beginning mark to determine the segmentation region for a context. The proposed adaptive segmentation method, which uses the feature map and a region growing method, can successfully segment the context region. The segmented area is compressed by a lossless coding algorithm, and the background area is compressed by a lossy coding algorithm.
Section snippets
Biologically motivated selective attention model
Fig. 1 shows the architecture of a selective attention model including the trainable Fuzzy ART network for inhibition and reinforcement [2]. In order to model the human-like visual bottom-up attention mechanism, we used the four bases of edge (E), intensity (I), color (RG and BY), and symmetry information (Sym). The roles of the retina cells and the lateral geniculate nucleus (LGN) are reflected in the proposed attention model [3]. The symmetry information is obtained by a noise-tolerant
Non-uniform image compression using the selective attention model
Fig. 2 shows a block diagram of the proposed non-uniform image compression using the bottom-up model and top–down trainable model. The bottom-up processing models a function of primitive selective attention in the human vision system since people selectively attend to a salient area according to various stimuli of an input scene. We can get a plausible area using the trainable top–down model through interaction with the supervisor. The most salient point is used as the beginning point to decide
Experimental results
We present the experimental results of the proposed non-uniform image compression using a human-like attention model. We use an input image with 320×240 pixels and 24 bit true color information in the experiments. Fig. 3 shows the simulation results of the proposed non-uniform image compression in a complex natural image. The evaluation of the proposed compression is measured by a compression ratio and a peak signal-to-noise ratio (PSNR). Table 1 shows the comparison with the uniform JPEG
Conclusion
We propose a non-uniform image compression method using a selective attention model for a natural image. The bottom-up model can identify a meaningful area and the top–down model can distinguish find a plausible object area from an unwanted salient area in a natural image. The proposed adaptive segmentation, using the feature map and a region-based segmentation, successfully segmented the context region from the whole image. Using the proposed method, the PSNR was enhanced with a slight
Acknowledgment
This research was funded by the Brain Science and Engineering Research Program of the Ministry of Korea Science.
References (11)
- et al.
Saliency map model with adaptive masking based on independent component analysis
Neurocomputing
(2002) - et al.
Seeded region growing
IEEE Transactions on Pattern Analysis and Machine Intelligence
(1994) - et al.
Biologically motivated visual attention system using bottom-up saliency map and top–down inhibition
Neural Information Processing—Letters and Reviews
(2004) Sensation and Perception
(1996)- et al.
Computational modeling of visual attention
Nature Reviews Neuroscience
(2001)
Cited by (5)
Weighted feature correlation and fusion saliency
2016, Communications in Computer and Information ScienceA NONUNIFORM HIGH-QUALITY IMAGE COMPRESSION METHOD to PRESERVE USER-SPECIFIED COMPRESSION RATIO
2011, International Journal of Image and GraphicsThe case for context-aware compression
2011, HotMobile 2011: The 12th Workshop on Mobile Computing Systems and ApplicationsA new efficient method for color image compression based on visual attention mechanism
2010, Proceedings of SPIE - The International Society for Optical EngineeringA non-uniform image compression using genetic algorithm
2008, Proceedings of IWSSIP 2008 - 15th International Conference on Systems, Signals and Image Processing