- You are required to install the newest Java Runtime Environment for executing the applet
- The Applet Version utilizes a Signed Appletwhich requires your trust for uploading your local image or accessing a remote URL image.
This applet demands a large memory. If you encounter java out of heap meamory, please increase your java runtime memory as follow:
- Open Java Control Panel:
From Windows Control Panel: Click “Start” at screen lower-left corner and choose Control Panel. Click “Java Plug-in”
.
From Java logo on Windows task bar (if a browser window has Applet plugin): Right-click Java logo and choose Open Control Panel - Select the Java tab and enter the Java Applet Runtime Settings -> View.
- In the box labeled “Java Runtime Parameters”, add (for memory limit 256MB) -Xmx256M
- Close Java Control Panel, and restart the browser if running.
Introduction:
Adaptive color quantization is a task to reduce the number of color in an image and is a significant process for various applications such as image compression and image segmentation.
However, most of the previous works are developed with the presumption that the number of color must fit the limit of devices. Instead, to pursue a high perceptive quality of images, a fast and flexible neural network approach based on growing quadtree structure self-organizing map (GQSOM) is designed. Unlike other conventional algorithms which always require the number of color to be determined at first, GQSOM is with the self-growing feature to generate color palettes of a fourth power number.
Kohonen’s self-organizing map (SOM) is an unsupervised and competitive learning neural network which is first used in Dekker’s work for color quantization and is shown higher quality than algorithms based on octree and median-cut. However, the cost is demanding much more computing time for training the neural network. The complexity of typical SOM algorithm is O(N), N is the number of neurons, caused from the bottleneck of searching best-match-unit (BMU) and further updating neighbor neurons as a winner-take-most (WTM) process. Apparently, as the number of output neuron increases, the time-consuming makes traditional SOM infeasible to be used.
This work makes eliminating psychovisual redundancy in a true-color image possible. Psychovisual redundancy is a phenomenon that certain information in an image is with less significance. Previous algorithms were designed for small quantization numbers , say, below 256 colors, to fit certain applications, but, in some natural color images, such 256 quantization color is not sufficient for representation. In other words, previous works did not take large quantization number into account. The proposed algorithm fixes the neuron number for searching BMU to make the search complexity becoming O(1). The price is that more memory space is required like a typical question of time-memory tradeoff. The program sets 45dB PSNR as a threshold of reducing psychovisual redundancy derived from an empirical study considering both object and subject criterion. A homogenous color wheel below is taken as a sample to simulate a critically quantized situation. In fact, in most of cases, 40dB PSNR is sufficient to produce a visually identical color image compared with the original one. You can run the demo program using the hyperlink of the wheel to observe the process.
