CAJ | 학술논문

通过建立一个多输出的BP神经网络，提取图像的底层特征作为网络的输入，用语义期望值作为网络的输出．训练完成后，该网络能够对风景图像进行多种语义分类检索，从而建立起了从底层特征到语义特征之间的映射．提出的一种颜色提取方法不仅降低了颜色特征向量的维数，减少了计算量，节省了时间，而且在描述风景图像的颜色内容上更加准确．如何选取图像的语义阈值是一个重点也是一个难点，通过实验发现，当阈值的选取范围在[0．55，0．65]时，检索的查全率和准确率能达到一个比较好的平衡效果．实验证明，此方法在风景图像的分类上取得了较好的检索查全率和准确率．
통과건립일개다수출적BP신경망락，제취도상적저층특정작위망락적수입，용어의기망치작위망락적수출．훈련완성후，해망락능구대풍경도상진행다충어의분류검색，종이건립기료종저층특정도어의특정지간적영사．제출적일충안색제취방법불부강저료안색특정향량적유수，감소료계산량，절성료시간，이차재묘술풍경도상적안색내용상경가준학．여하선취도상적어의역치시일개중점야시일개난점，통과실험발현，당역치적선취범위재[0．55，0．65]시，검색적사전솔화준학솔능체도일개비교호적평형효과．실험증명，차방법재풍경도상적분류상취득료교호적검색사전솔화준학솔．
This paper establishes a multioutput BP neural network. This method extracts lowlevel fea tures vector as the network input and the exceptions as its output. The System trains the network with BP arithmetic. When the training is over, this network can classify natural images. So it has established the mapping between the lowlevel features and highlevel semantic features. This paper proposes a new Color method which not only reduces the color characteristic vector dimension and the amount of computation＇, saves the time, but also describes the content of image more accurately. How to select the semantic image threshold is important and difficult. Through the experiment, it is found that when the selected threshold ranges of[0.55, 0.65], the retrieval of the recall rate and the accurate rate can achieve a better balance effect. The experiment proves that it has obtained the high accuracy.