中国科技论文
中國科技論文
중국과기논문
Sciencepaper Online
2014年
7期
803-807,811
,共6页
图像%分类%图像块%特征
圖像%分類%圖像塊%特徵
도상%분류%도상괴%특정
images%classification%patch%feature
针对以关键字进行检索分类的图像,首先利用经验设置规则提取初始图像块;对于类中每个备选图像块,寻找所有备选图像块中与其最相似的图像块,并根据梯度方向直方图特征进行聚类,形成特征组;最后,利用图像注册的方式将特征组进行合并,形成商品图像完整的特征。本文以电商图像为例对算法进行了测试,实验结果表明,辨识性特征区域提取的准确率可达70%以上,且与已有方法相比提取特征区域更加完整。
針對以關鍵字進行檢索分類的圖像,首先利用經驗設置規則提取初始圖像塊;對于類中每箇備選圖像塊,尋找所有備選圖像塊中與其最相似的圖像塊,併根據梯度方嚮直方圖特徵進行聚類,形成特徵組;最後,利用圖像註冊的方式將特徵組進行閤併,形成商品圖像完整的特徵。本文以電商圖像為例對算法進行瞭測試,實驗結果錶明,辨識性特徵區域提取的準確率可達70%以上,且與已有方法相比提取特徵區域更加完整。
침대이관건자진행검색분류적도상,수선이용경험설치규칙제취초시도상괴;대우류중매개비선도상괴,심조소유비선도상괴중여기최상사적도상괴,병근거제도방향직방도특정진행취류,형성특정조;최후,이용도상주책적방식장특정조진행합병,형성상품도상완정적특정。본문이전상도상위례대산법진행료측시,실험결과표명,변식성특정구역제취적준학솔가체70%이상,차여이유방법상비제취특정구역경가완정。
An image feature identification method is proposed to classify the images according to the key words.Firstly,an empir-ical rule is set to achieve initial patches.Then,the nearest patches of every initial patch are selected.The features of HOG(histo-grams of oriented gradient)are employed to cluster them as the feature group.Finally,the patches are combined to complete fea-ture by the algorithm based on image registering.In the experiments,proposed methods are implemented to analyze E-commerce images.The accuracy of identifiable feature extraction is as high as 70%,and the extracted feature region is more complete than current methods.