Research On Detection Of Image Including Regions Of Interest Based On Pcnn And Its Applications

Pulse Coupled Neural Network (PCNN) is a new method of image processing proposed in recent years, and it has been applied in varied image processing domains such as noise removal, image segmentation, image shadow removal, etc. Applying PCNN to detecting regions of interest (ROI), the results of image enhancement based on PCNN which can stand out ROI and wipe off unrelated regions, can reduce range of searching ROI effectively for ROI localization algorithm, which can improve the accuracy of localization ROI and processing efficiency of these algorithms. However, the results of PCNN enhancement are serial images, and it is issue of applying PCNN to solve practical problems how to achieve automatically the image including ROI from the serial images. Research on detection of ROI has not only theoretical value, but also social and economic significance.This paper focus on study on detection of image including ROI based on PCNN and its applications. The main content and novelties of this thesis are summarized as follows:(1) Being aimed at the deficiencies of standard and improved PCNN, a modified PCNN is proposed which is named as LMO-PCNN, and two sub-models can be obtained by different output prescripts of LMO-PCNN, they are named as LAO-PCNN and LIO-PCNN. At the same time, combined with network parameters, dynamic behaviors of LMO-PCNN are analyzed theoretically and deduced by formula in detail. The improved model not only maintains the basic characteristics of PCNN, but also has some new features.(2) Apply the two sub-model of LMO-PCNN to different applications of image processing. Combine the LAO-PCNN and median filter to recover mixed-noise contaminated images. And the optimal methods of setting key parameters are achieved by mathematical reasoning and experiments. An enhancement method of image including ROI is proposed based on LIO-PCNN. And parameters of LIO-PCNN are analyzed theoretically and deduced by formula in detail. By different parameters setting, two methods of image enhancement based on LIO-PCNN can be obtained, named as elicitation method and adaptive method. Application advantages of algorithms are analyzed. Experimental results show that the new algorithms improve processing performance obviously than current algorithm. (3) To improve the accuracy of localization ROI and processing efficiency of algorithms, considering of reducing range of searching ROI, combined with Independent Component Analysis (ICA), an algorithm of detecting image including ROI automatically is proposed based on LIO-PCNN, which can detect the image result including ROI automatically from image serial obtained by LIO-PCNN image enhancement, and this result can stand out ROI and wipe off unrelated regions efficiently. The average runtime of the detection method proposed in this paper is 0.08s, and the accuracy of target image outputs is 98.6%. At the same time, it is robust to shape and position of ROI.(4) Utilizing the characteristic of car plate in the result images obtained from performance of detection method based on LIO-PCNN-ICA, combined with region identification, a method of localizing car plate is proposed, and the algorithm can locate the car plate precisely and segment the characters simultaneously. Experimental results show that using LIO-PCNN-ICA can optimize image input of localization algorithm and reduce range of searching car plate region to achieve higher positioning accuracy and faster implementation of efficiency than current algorithm. This also shows that LIO-PCNN-ICA has the advantages of improving the positioning accuracy and processing efficiency of the follow-up ROI location algorithm.(5) Based on point of view ROI detection, a method based on LMO-PCNN for Raman spectra qualitative analysis was proposed. Firstly, LMO-PCNN is improved, and combined with network parameters, dynamic behaviors of LMO-PCNN are analyzed theoretically. Then, after encoding the Raman spectra by using improved LMO-PCNN neurons'characteristics of fatigue and refractory period, the improved algorithm was used to match the code corresponding to the detected sample with all of the base code in the database one by one, and then their matching similarity were acquired to determine the sample type. Meanwhile, traditional qualitative analysis method based on spectral template has some deficiencies, like that it is difficult to determine the characteristic peak of the detected sample and the matching analysis process has a high degree of redundancy. While our proposed method not only can avoid these deficiencies very well, but also need a small amount for data storage, the requirement of the matching degree and the storage space was only 1/3 and 5.8% of it used in the traditional qualitative analysis method based on spectral template.