Research On PCNN Model Improvement And Parameters Adjustment

Pulse Coupled Neural Network(abbreviated as PCNN) is a new artificial neural network model following the networks whose computing unit is based on perceptions and activation function. As considering the characteristics of neurobiology of time coding and spatial accumulation, PCNN get incomparable advantages comparing with the traditional artificial neural network including pulse coupling, nonlinear multiplication modulation,pulse capture in neighborhood neurons and exponential decay threshold mechanism, etc.Therefore, PCNN has broad application prospects in image segmentation.There is still a certain distance between PCNN model and real biological vision model and a room for further development and refinement notwithstanding research on PCNN has deepened successively in recent years. Parameters adjustment of the model,excavating its characteristics and the improvement of its efficiency gradually become major focuses.Firstly, PCNN parameters are combined with three important image local features that are image gradient, local image entropy and gray correlation degree to achieve linking strength self-adjusting systems and link weight matrix self-adjusting systems. Then, the best iteration number and the results of image segmentation are analyzed depending on different criteria.Secondly, considering some regions in the segmentation result using non-linking PCNN are inconsistent with the human visual judgment of target and background and the capture characteristics of pulse-coupled neuron are single-direction, fully pulse capture characteristics are used to locate the dark regions that met the definition which is described as high similarity region in object areas and inversed fully pulse capture characteristics are used to locate the light regions that met the definition which is described as high similarity region in background areas. Then, the segmentation result is obtained by executing logical operations between the segmentation result using non-linking PCNN and high similarity regions. Extensive experimental evaluation demonstrates that this method maintains the integrity of the segmentation result.Thirdly, an improved model is developed by introducing synaptic integration in neurobiology and fully pulse capture characteristics into PCNN to overcome the inherent defects in image segmentation. Extensive experimental evaluation demonstrates that theproposed model achieves good performance in image segmentation and the objective results are consistent with the subjective results.