High-tech and Artificial Intelligence Make the Invisible Visible

Research and Development in Biosystems Engineering

Content Adaptive Image Compression

High throughput screening with imaging systems is growing increasingly important in biomedical applications and material sciences too. Extremely large quantities of image data are generated and the images are usually compressed to save memory. There are universal methods, chiefly transformations, suited for rapid (on-the-fly) compression (e.g. cosine transforms in JPEG), and slower (less data loss or higher compression) methods (wavelet transform contained in JPEG2000) used more for archiving. Each of these methods, however, treats every image separately and de facto independently of its content. Taking this as the point of departure, compression can also be executed as a function of the typical content of an image from a larger set of images. A number of artificial neural networks are trained for specific contents of images and their compression. Investing effort in this one-time training is worthwhile whenever a large number of similar images will be saved later, e.g. images (microscopy, CT, NMR) of an identical tissue or organ. This is often the case in biological and medical applications. The advantage over universal methods is the higher compression performance through the a priori information on the dominant image content incorporated in the compression. Comparisons with similar image formats (e.g. JPEG, JPEG2000, PNG) as well as universal compression algorithms (e.g. ZIP) have demonstrated this method’s superiority. The method can be broadly parameterized – from nearly lossless to extremely high compression rates. By selecting typical image contents to train the compressors, the quality of reconstructed images to also be rendered as a function of content, i.e. relevant structures are compressed less and thus with less loss than other regions of images. This feature is independent of the local appearance of independent structures in an image. The integration of this system in commercial image acquisition and/or storage systems is planned for the near future.

Systemdiagram der inhaltsadaptiven Bildkompression
© OVGU

Schematic of the content adaptive image compression systems. First, image content is analyzed as an image block (top). Based on the image content, every image block is processed by a prepared compressor. The separate compressors are configured as an autoassociative feed-forward network (left).

Systemdiagramm zur inhaltadaptiven Bildkompression
© OVGU

The individual compressors are trained only with the image blocks of the related class (cluster). Images are compressed by adjusting the middle layer’s width (number of neurons). This system received the Saxony-Anhalt Innovation Award in 2006.

© Medizinische Hochschule Hannover

Medical diagnostics produce millions of similar images that must be archived for the purpose of documentation. Our system is ideal for such needs. The extensive data set from which this image was taken is from Hannover Medical School.

Publications

Seiffert, U.: Content Adaptive Compression of Images Using Neural Maps. Proc. of the International Workshop on Self-Organizing Maps WSOM 2005, Paris, France, pp. 227-234, 2005.

Seiffert, U.: ANNIE - Artificial Neural Network-based Image Encoder, Neurocomputing, 2014, 125, 229-235