Research On Near-duplicate Image Detection Algorithm Based On Deep Learning Features

The existing deep learning-based near-duplicate image detection algorithms mostly rely on the CNN(Convolutional neural network)features.Compared with the traditional handcrafted features,some CNN-based detection algorithms have achieved a great improvement.However,these CNN-based detection algorithms only extract global CNN features or local CNN features for near-duplicate image detection,without fully considering the respective advantages of global CNN features and local CNN features.Therefore,there is still a lot of room for improvement in the performance of near-duplicate image detection.Moreover,since existing methods neglect the essential differences between image copies and similar images,it is difficult for them to identify image copies from similar images.To solve the above problems,this paper proposes two near-duplicate image detection algorithms based on CNN features.(1)A near-duplicate image detection algorithm based on global and local CNN features.In order to exploit the advantages of both global CNN feature and local CNN features,this paper proposes a near-duplicate image detection algorithm based on global and local CNN features.The proposed algorithm mainly consists of a coarse matching stage and a fine matching stage.In the coarse matching stage,the proposed algorithm extracts and matches global CNN features between a given query image and database images to quickly filter out most irrelevant images of the query.And in the fine matching stage,a novel local feature extraction scheme is proposed to further improve the detection accuracy.In this scheme,CFMs(Convolutional Feature Maps)are firstly generated by feeding images to a CNN model.Then,this scheme detects the points with maximum activations on these generated CFMs.Finally,the scheme selects the patches surrounding the detected points as the local regions.In addition,the algorithm also introduces attention mechanism in the fine matching stage.By detecting visual saliency maps,the proposed algorithm obtains the local patches of interest to human visual system as the supplement of the above local regions.By exploiting the advantages of both global and local CNN features,the proposed algorithm can achieve real-time and accurate nearduplicate image detection.(2)A near-duplicate image detection algorithm based on residual domain-based deep learning features.In order to accurately identify image copies from similar images,this paper proposes a nearduplicate image detection algorithm based on residual domain-based deep learning features.The proposed algorithm first extracts and matches SIFT(Scale Invariant Feature Transformation)features between a given query image and test image.Then,the method makes the use of properties(including characteristic scale,dominant orientation and coordinates)of SIFT features to align the query image to the test images.The overlapping regions are then extracted from the aligned query image and the test image,respectively.Finally,the residual images,which are fed to an optimized CNN model for learning the copy relationship between the two images,are generated by subtracting the geometrically aligned original image from the test images.The extensive experiments demonstrate that the proposed algorithm can accurately identify image copies from similar images,and achieve state-of-the-art performances in the aspects of detection accuracy and training efficiency.