本文選題：特征連接模型 + 圖像增強��； 參考：《蘭州大學》2017年碩士論文

【摘要】：脈沖耦合神經(jīng)網(wǎng)絡(luò)(Pulsed Coupled Neural Networks,PCNN)是第三代神經(jīng)網(wǎng)絡(luò)的典型代表,演化自哺乳動物視覺皮層系統(tǒng)的同步脈沖發(fā)放現(xiàn)象.在研究PCNN的基礎(chǔ)上,我們提出了特征連接模型(Feature Linking Model,FLM),利用FLM的賦時矩陣和單通工作方式進行圖像處理.首先,我們提出了FLM,該模型有反饋輸入和連接輸入兩個輸入端,它與PCNN有相似的結(jié)構(gòu),但是在PCNN中有三個漏電積分器而FLM只有兩個漏電積分器,因此FLM較PCNN簡單.我們發(fā)現(xiàn),當閾值呈現(xiàn)指數(shù)衰減時,FLM的賦時矩陣和刺激輸入之間呈現(xiàn)一個對數(shù)關(guān)系,并且通過單通工作方式記錄了脈沖發(fā)生的時間.FLM中的全局抑制項,提高了同一個區(qū)域神經(jīng)元的同步性和不同區(qū)域神經(jīng)元之間的異步性.另外,γ帶振蕩啟發(fā)的連接調(diào)節(jié)機制和動態(tài)閾值特性,使得FLM更接近于生物神經(jīng)元特性.此外,受到生物神經(jīng)學支持的賦時矩陣也是本文研究的重點.其次,我們提出了FLM的單通工作方式,該工作方式可以使得所有的神經(jīng)元只能點火一次,且保證所有的神經(jīng)元都能夠點火.我們利用FLM的單通工作方式獲得了賦時矩陣,這為本文提出的圖像處理算法提供了基礎(chǔ).此外,FLM是通過其兩種突觸輸入來獲得同步脈沖的,我們分別介紹了這兩種突觸的調(diào)節(jié)機制和這兩種突觸的波形傳播形式.基于FLM的賦時矩陣,結(jié)合同步特性,我們提出了FLM圖像增強、圖像分割和圖像復原三種方法,并詳細介紹了每種方法的預(yù)處理操作、算法的具體實現(xiàn)過程、參數(shù)的設(shè)定原理和與其它方法的比較實驗等,最后我們從主觀和客觀兩個方面對本文的算法進行評價,評價結(jié)果證明了本文提出的算法的優(yōu)越性.
[Abstract]:Pulse coupled neural network (PCNN) is a typical representation of the third generation neural network, which evolves from the phenomenon of synchronous pulse firing in mammalian visual cortex system. Based on the study of PCNN, we propose a feature linking Model (FLM), which uses the time matrix of FLM and the single way to process the image. Firstly, we propose FLM, which has two input terminals: feedback input and connection input. It has a similar structure to PCNN, but there are three leakage integrators in PCNN and only two leakage integrators in FLM, so FLM is simpler than PCNN. We find that there is a logarithmic relationship between the time matrix of the FLM and the stimulus input when the threshold is exponentially attenuated, and the global suppression term in the pulse generation time. FLM is recorded by a single way. The synchronization of neurons in the same region and the asynchronism among neurons in different regions are improved. In addition, the connection regulation mechanism and the dynamic threshold characteristics of 緯 -band oscillation elicited the FLM to be closer to the biological neuron characteristics. In addition, the time matrix supported by biological neurology is also the focus of this paper. Secondly, we propose a single way of FLM, which can make all neurons light fire only once, and ensure that all neurons can light fire. We obtain the time matrix by using the single way of FLM, which provides the basis for the image processing algorithm proposed in this paper. In addition, the two synaptic inputs of FLM are used to obtain the synchronous pulse. We introduce the modulation mechanism of these two synapses and the waveform transmission forms of these two synapses, respectively. Based on the timed matrix of FLM and the synchronization characteristics, we propose three methods of FLM image enhancement, image segmentation and image restoration, and introduce the preprocessing operation of each method and the implementation process of the algorithm in detail. Finally, we evaluate the algorithm from the subjective and objective aspects, and the evaluation results prove the superiority of the proposed algorithm.
【學位授予單位】：蘭州大學
【學位級別】：碩士
【學位授予年份】：2017
【分類號】：TP391.41;TP183

【參考文獻】

相關(guān)期刊論文 前1條

1 章毓晉;中國圖像工程及當前的幾個研究熱點[J];計算機輔助設(shè)計與圖形學學報;2002年06期

相關(guān)碩士學位論文 前1條

1 高春霞;基于脈沖耦合神經(jīng)網(wǎng)絡(luò)和進化算法的圖像分割方法研究[D];西安電子科技大學;2007年

，

本文編號：2066965