高技术通讯
高技術通訊
고기술통신
HIGH TECHNOLOGY LETTERS
2014年
10期
52-64
,共13页
谢康%杨义先%张玲%杜晓峰%辛阳
謝康%楊義先%張玲%杜曉峰%辛暘
사강%양의선%장령%두효봉%신양
多车牌%边缘检测%细胞神经网络(CNN)%同心邻域极值(CNE)%粒子群优化(PSO)
多車牌%邊緣檢測%細胞神經網絡(CNN)%同心鄰域極值(CNE)%粒子群優化(PSO)
다차패%변연검측%세포신경망락(CNN)%동심린역겁치(CNE)%입자군우화(PSO)
multi license plate%edge detection%cellular neural network (CNN)%concentric neighborhood extreme (CNE) value%particle swarm optimization (PSO)
针对现有智能交通系统（ITS）多车牌定位识别算法漏检率高、处理速度慢等问题，在研究细胞神经网络（CNN）理论的基础上，提出了一种基于CNN同心邻域极值（CNE）的ITS图像多车牌区域边缘检测算法，简称CNECNN算法。该算法只需计算CNN中同心邻域内极大值与极小值函数差的二阶微分零交叉点，即可获得图像的边缘。此外，该算法利用CNN稳态能量函数惩罚约束机制优化粒子群适应度函数，在解空间中搜索参数全局最优解以获得CNN邻域极值模板参数。该算法为并行算法，具有运算量小，易于大规模集成电路实现，能够克服早熟收敛等优点。实验结果表明，与传统边缘检测算子和CNN通用机（CNNUM）固定模板参数算法相比，该算法漏检度降低了12.9%。
針對現有智能交通繫統（ITS）多車牌定位識彆算法漏檢率高、處理速度慢等問題，在研究細胞神經網絡（CNN）理論的基礎上，提齣瞭一種基于CNN同心鄰域極值（CNE）的ITS圖像多車牌區域邊緣檢測算法，簡稱CNECNN算法。該算法隻需計算CNN中同心鄰域內極大值與極小值函數差的二階微分零交扠點，即可穫得圖像的邊緣。此外，該算法利用CNN穩態能量函數懲罰約束機製優化粒子群適應度函數，在解空間中搜索參數全跼最優解以穫得CNN鄰域極值模闆參數。該算法為併行算法，具有運算量小，易于大規模集成電路實現，能夠剋服早熟收斂等優點。實驗結果錶明，與傳統邊緣檢測算子和CNN通用機（CNNUM）固定模闆參數算法相比，該算法漏檢度降低瞭12.9%。
침대현유지능교통계통（ITS）다차패정위식별산법루검솔고、처리속도만등문제，재연구세포신경망락（CNN）이론적기출상，제출료일충기우CNN동심린역겁치（CNE）적ITS도상다차패구역변연검측산법，간칭CNECNN산법。해산법지수계산CNN중동심린역내겁대치여겁소치함수차적이계미분령교차점，즉가획득도상적변연。차외，해산법이용CNN은태능량함수징벌약속궤제우화입자군괄응도함수，재해공간중수색삼수전국최우해이획득CNN린역겁치모판삼수。해산법위병행산법，구유운산량소，역우대규모집성전로실현，능구극복조숙수렴등우점。실험결과표명，여전통변연검측산자화CNN통용궤（CNNUM）고정모판삼수산법상비，해산법루검도강저료12.9%。
The cellular neural network (CNN) theory was applied to the study of edge detection in the multi license area of an intelligent transportation system (ITS)’ images to improve the ITS’ performance in license plate recognition, and a new edge detection algorithm based on the concentric neighborhood extreme (CNE) value of CNN, called CNECNN algorithm for short, was put forward. To get the edge area, this algorithm calculates the zero crossing point of a difference function that depends only on the concentric neighborhood extreme value. In order to obtain the CNN template parameters, an energy function constraint method is used to construct a new fitness function of particle swarm optimization (PSO), jumping out the premature convergence, and ultimately find the optimal solution. This new approach can be easily used for VLSI implementation because of its parallelism. Compared with the traditional edge detection operators and general edge detection template in CNN Universal Machine (CNNUM), the simulation results of the images collected in the real environment show that the algorithm based on CNECNN can reduce the miss rate by 12.9%.
