Pulse Coupled Neural Network And Its Applied Research And Face Detection, Image Enhancement

Pulse Coupled Neural Network (PCNN) which shows bio-visual characteristics is a new generation of artificial neural network. Research on PCNN has important and significance value in theories and applications. Based on the research on the therotical model of traditional PCNN and its characteristics of pulse parallel propogation, the dissertation improves traditional PCNN model in terms of specific applications on image processing and analysis, such as image segmentation, extreme impulse noise filtering, noise and speckle suppression in ultrasound Dopplar blood flow spectrograms and face detection under natural background.The main contributions of this dissertation are as follows:1. A novel PCNN model based on grayscale iteration threshold (GIT-PCNN) is first brought forward. GIT-PCNN simplifies traditional PCNN model and reduces the related parameters. GIT-PCNN improves the threshold of traditional PCNN into a grayscale iterative threshold which is related to the grayscale statistics of the original image. When GIT-PCNN is used for image segmentation purpose there is no iteration or iteration end criteria required. Experimental results show that GIT-PCNN model improves the segmentation speed while keeps the segmentation performance undegraded when compared to existing PCNN segmentation algorithms based on maximum Shannon entropy. In addition, GIT-PCNN achieves segmentation performance and faster than the PCNN segmentation algorithm based on block preprocessing, maximum Shannon entropy and minimum cross-entropy.2. A simplified PCNN model, named as single-linking PCNN is proposed as extreme impulse noise filter. The single-linking PCNN model simplifies traditional PCNN model, therefore there is no parameter selection or iteration is required when the model is applied as a filter. The threshold of the model and the filtering time can be adaptively determined based on the characteristics of extreme impulse noise. Experimental results show that the proposed filter algorithm demonstrates better performance than classical impulse noise filtering algorithms, such as traditional median filter, standard median filter, convolution-based impulse detector and switching median filter and decision-based filtering algorithm when noise intensity varies from 10% to 60%. Meanwhile, the proposed algorithm reaches better filtering performance than existing PCNN filter for extreme impulse noise when noise intensity is up to 80%.3. A two-step processing model, MP-PCNN (Matching Purisuit with PCNN) is proposed for removing noise and speckles in the STFT (Short Time Fourier Transform) spectrogram of Doppler ultrasound blood flow signals. MP is first applied to decompose the signal into a set of linear atoms based on a Gabor dictionary and then the atoms closely related to the dictionary are defined as signal and the residue is defined as noise. After removing background noise from the signal, a uni-directional threshold decaying PCNN model is brought forward to detect speckles in the STFT spectrograms. The proposed PCNN model simplifies traditional PCNN model so that the parameters can be applied on various STFT spectrograms for speckle detection and removal. Experimental results show that the MP-PCNN has better performance than the classical MPWD (Matching Pursuit with Wigner Distribution) algorithm for removing noise and speckles in the STFT spectrograms of ultrasound Dopplar blood flow signal.4. A face detection algorithm based on unit-linking PCNN time signature is proposed, in which skin block detection is applied to improve the speed of the algorithm. The number of face contained in a test image is determined by clustering the upper-left corner coordinates of the detected face blocks. Experimental results show that the proposed method can detect faces in a test image under natural background even when the test image is noised, rotated or scaled.