CNN特征与BOF相融合的水下目标识别算法

为了改善作为低级表示的尺度不变特征变换(scale invariant feature transform, SIFT)匹配常出现的没有足够特征来防止假匹配的问题,提出在传统方法“词袋”(bag of features, BOF)算法中融合具有较好语义分割能力的卷积神经网络(convolution neural network, CNN)特征来提高识别率的方法。利用ImageCLEF网站的LifeCLEF鱼类视频,制作目标图像数据库。在caffe平台的Alexnet模型进行卷积神经网络的训练,提取图像库和查询图像的特征。利用训练好的CNN特征在Matlab软件进行识别试验验证,计算汉明距离来验证匹配效果。改变参数值来观察不同汉明距离阈值对水下目标识别结果的影响。自制图像库的试验表明,融合深度学习的特征可以有效提高BOF算法的水下目标识别率,对汉明距离阈值的选择需要根据实际情况选择合适的参数。

Algorithm of underwater target recognition based on CNN features with BOF

In order to prevent false matching problems of scale invariant feature transform (SIFT) matching as a low-level representation for lack of sufficient features, an improved bag of features (BOF) algorithm method combined with the convolution neural network (CNN) features was proposed, which had better semantic segmentation ability to enhance the recognition rates. The LifeCLEF fish video on ImageCLEF website was used to create our own target image databases. Convolution neural network was trained in the Alexnet architecture of caffe, and the features of image databases and query images were extracted. The trained CNN features were simulated in Matlab, and the hamming distance was calculated to verify the matching effect. In addition, the parameter values were changed to test the effect of different Hamming distance thresholds on target recognition results. The experiment of self-made image databases showed that the fusion of depth learning features could effectively improve the underwater target recognition rates of BOF algorithm, and the selection of Hamming distance thresholds required selecting the appropriate parameters according to the actual situation.