Texture Synthesis And Facial Feature Detection Using DTCWT

As a multi-resolution representation of signals, the wavelet transform has been a powerful tool in image and graphics processing. Unlike the Fourier transform, the wavelet transform has locality and scalability in both spatial and frequency domain. This feature makes the wavelet particularly suitable for signals with both smooth area and high discontinuities, which is often the case for images and graphics. However, the performance of traditional separable 2D wavelets is impaired by its lack of directionality. In this thesis, a new type of directional wavelet transform, the dual-tree complex wavelet transform (DTCWT), is introduced. Recently designed by Kingsbury, the DTCWT can capture different directional features in the image with different subbands, while preserving most nice features of the standard discrete wavelet transform. In the past decade, a lot of research has been dedicated in its applications on image denoising, compression, etc. This thesis mainly focuses on the application of DTCWT on texture synthesis and facial feature detection.Texture is an important feature in images, and texture synthesis has a variety of applications in MPEG compression, image recovery, filter design and so on. Based on previous research, we propose a DTCWT-based texture synthesis algorithm. Instead of generating new textures in the spatial domain, we first transform the sample texture image into a set of six directional wavelet subbands, and small patches are selected from the sample subbands to form new synthesized subbands according to a similarity evaluation function. In the end, the synthesized texture image is generated from the new subbands by the inverse transform. Compared to other existing methods, our algorithm has higher efficiency and better performance, especially for structured textures.Human face is also one of the most important visual features in images and videos. Face detection and facial feature detection is a key problem in facial information processing and content-based image indexing. We propose a facial feature detection algorithm by taking advantage of the orientation selective feature of DTCWT. Local and global energy functions are computed by combining coefficients from different directional subbands. Feature points are extracted by searching for local maxima of the energy function. Examples show that the proposed algorithm has good performance. With no need of training, it is also faster than other algorithms.A DTCWT processing platform system is also designed for more flexible application of the DTCWT. It contains a preprocessing module, a DTCWT transform module and an application module. Both internal and external interfaces are provided.