Color Image Compression Based On Quaternion Neural Network

In this thesis, color image compression technique based on quaternion theory is studied, including quaternion time-frequency characteristics image compression, quaternion neural network principal components image compression and quaternion time-frequency pretreatment principal components image compression. The color image can be described as a quaternion matrix, using this quaternion matrix as coding object data, making use of algorithms such as quaternion discrete cosine transform, quaternion neural network principal analysis, etc. to extract the principal components, after further operations related to operation of the quaternion, the image data compression is realized.The emphases in this thesis are summarized as follows:Firstly, quaternion discrete cosine transform (QDCT) is introduced, also image compression using QDCT is proposed, while QDCT can assemble the image information at low frequencies, then by using low-frequency filters to get the low frequencies to reconstruct the image. Lastly, the image compression is realized. The proposed algorithm's efficiency is simulated by the software, and the application significance of the project is discussed. This is an innovation as well as a try of the thesis.Secondly, quaternion principal components analysis is introduced, then quaternion principal components analysis based neural network is proposed, adopting the quaternion generalized Hebbian algorithm, the obtained weight can successfully extract the principal components of the input data, using this weight can realize the color image compression and reconstruct. The algorithm is also simulated by the software. The results show it is more efficient than quaternion principal components analysis. The obtained weight has a good ability of generalization, can be used to compress and reconstruct on other images. This is the main innovation of the thesis.Finaly, combining with the image pretreatment technique application characteristics of the normal number fields, the color image compression using QDCT pretreatment method to quaternion neural network principal components analysis is proposed. Firstly, color image data using QDCT to pretreatment work, then with the neural network principal components analysis to the transformed data. This way can accelerate the speed of the network's learning rate and get higher image compression quality, the obtained weight also has a good ability of generalization, can be used to compress and reconstruct on other images. This is another innovation of the thesis.