Compression of a Signature Database
MetadataShow full metadata
Lossy compression algorithms are commonly used to compress multimedia data such as image, video, and audio. This type of compression method is generally applied in streaming media applications since there is a strong motivation for reducing the storage space and/or transmission bandwidth where small visual quality loss is acceptable.
This thesis investigates and evaluates lossless and lossy compression algorithms via the compression and reconstruction of an object representation referred to as signature. Lossy compression is investigated due to the fact that lossless compression does not yield significant compression. An object signature is a sequence of values representing the distance between the object center and its boundary.
Several steps are involved in extracting an object signature. The processes applied to an input image are executed in the following order: binarization, connected component labeling (CCL), contour tracing, and signature calculation. This sequence of events is applied to a set of synthetic images. The output results are used to create an image signature database.
The lossy compression methods studied in this project are Differential Pulse Code Modulation (DPCM), Differential Linear Predictive Coding (DLPC), Discrete Cosine Transformation (DCT), and Compressed Sensing (CS). The lossless method used in this thesis is the Dictionary based compression method utilized by the UNIX 'gzip' compression utility. These algorithms are applied to the object signatures before storing them in a database. The compression quality is evaluated with respect to three different aspects: object recognition using a compressed signature, Signal-to-Noise ratio (SNR) of the original signature compared to the reconstructed compressed signature, and pixel by pixel comparison between the original image and the image reconstructed from a compressed signature.
Our results show that the recognition process is able to identify all input objects for every image and compression method combination. The experiments demonstrate that lossless compression is not a viable method for this application. Furthermore, according to the experimental results DPCM has the best rate distortion performance. On the other hand, the Compressed Sensing method produces higher distortion and requires a higher threshold value in order to accept a valid recognition response. This higher distortion caused by Compressed Sensing is also visible in the lower SNR values obtained when comparing the signature error with other compression methods utilized in this project. Although the quality results for this method are not as good as the DPCM method, Compressed Sensing has an important advantage of requiring less data to represent an image during the acquisition process. This means that CS might enable to capture the data with a smaller number of sensors.