光子学报
光子學報
광자학보
ACTA PHOTONICA SINICA
2009年
10期
2712-2716
,共5页
李念永%梁艳梅%张舒%杨立%常胜江
李唸永%樑豔梅%張舒%楊立%常勝江
리념영%량염매%장서%양립%상성강
文本定位%神经网络%小波变换%彩色图像
文本定位%神經網絡%小波變換%綵色圖像
문본정위%신경망락%소파변환%채색도상
Text location%Neural network%Wavelet transform%Color images
针对复杂彩色图像中文本的特征,提出了基于小波变换和BP神经网络甄别文本区域的算法.该算法首先利用文本块的边缘特征遴选出备选图像块,而后采用小波变换提取备选图像块的纹理特征,把这些纹理特征参量连同图像块的颜色特征和笔画特征参量输入训练好的BP神经网络,判断备选图像块是否包含文本.该方法运算简单,定位时间短.采用专用的文本定位比赛用图进行实验的结果表明,定位准确率可达到92%,召回率为87.4%.
針對複雜綵色圖像中文本的特徵,提齣瞭基于小波變換和BP神經網絡甄彆文本區域的算法.該算法首先利用文本塊的邊緣特徵遴選齣備選圖像塊,而後採用小波變換提取備選圖像塊的紋理特徵,把這些紋理特徵參量連同圖像塊的顏色特徵和筆畫特徵參量輸入訓練好的BP神經網絡,判斷備選圖像塊是否包含文本.該方法運算簡單,定位時間短.採用專用的文本定位比賽用圖進行實驗的結果錶明,定位準確率可達到92%,召迴率為87.4%.
침대복잡채색도상중문본적특정,제출료기우소파변환화BP신경망락견별문본구역적산법.해산법수선이용문본괴적변연특정린선출비선도상괴,이후채용소파변환제취비선도상괴적문리특정,파저사문리특정삼량련동도상괴적안색특정화필화특정삼량수입훈련호적BP신경망락,판단비선도상괴시부포함문본.해방법운산간단,정위시간단.채용전용적문본정위비새용도진행실험적결과표명,정위준학솔가체도92%,소회솔위87.4%.
According to the features of text information in complex color images,a method which is used to discriminate the text region on the basis of wavelet transform and BP neural network is proposed. The wavelet transform is adopted to extract the texture feature parameters of candidate image blocks,which is obtained by the judgment of the edge feature. Then these parameters are taken as input for the BP neural network which has been well trained, together with the parameters of the color feature and the stroke feature,to judge whether some texts are in the candidate image blocks or not. Compared with the others, the method mentioned above is simpler in operation and shorter in time of location. The experimental results indicate that, as for the images which are provided by the text location competition, locating accuracy can come to 92%,meanwhile, the recall ratio can reach up to 87.4%.