结合分类卷积神经网络和形状上下文的线画图检索

Line Drawing Retrieval Combining Classification CNN and Shape Context

传统线画图检索中仅仅利用线画图形状特征信息导致检索准确率不高,为了高效、准确地从线画图数据集中检索相似的线画图,提出一种结合分类卷积神经网络和形状上下文的线画图检索方法.首先利用大卷积核的分类卷积神经网络对线画图像数据集进行分类训练任务得到神经网络权值参数,使用该网络结构提取数据集中每张线画图的卷积特征信息;然后根据用户在画图板上绘制得到的简单线画图输入,利用卷积神经网络进行二次分类得到前15种最相似的分类,并结合形状上下文算法对15种分类匹配相似度并取前8种分类;最后使用卷积神经网络提取用户输入的线画图特征信息并与8种分类中的线画图特征信息进行匹配,根据相似度大小排序得到线画图匹配结果.基于Caffe卷积神经网络开发框架,采用TU-Berlinsketchbenchmark线画图数据集进行实验的结果表明,该方法能高效、准确地从数据集中检索得到相似线画图,同时能保证检索结果集中于最相似的几种类别且同类型中能有更多的选择.

Traditional line drawing retrieval technique is usually inaccurate or inefficient due to its dependent on user-defined shape features. To address these issues, this paper presents a novel line drawing retrieval method combined with classification convolution neural network(CNN) and shape context. First, a convolution neural network with large volume kernel is introduced, and the network weights can be trained by the classification task of the line drawing image dataset. The convolution feature information of each line drawing is also extracted with the network structure. Then, according to the input simple line drawing that the user draws on the drawing board, the top 15 kinds of classification are obtained by the two-step classification using CNN, and the similarity of 15 classes is matched with the shape context from which the top 8 classes can be chosen. Finally, the convolution neural network is used to extract the feature information of the input line drawing and match with the top 8 classes using their similarity measure. Experimental results show that our method can efficiently and accurately retrieve the similar line drawings from the TU-Berlin sketch benchmark dataset under Caffe framework, which ensure that the retrieval results can be concentrated in the most similar image category and the same type has more choice to be selected.