Application of multi-step image segmentation for near-duplicate image recognition

Abstract

The urgency of the paper is caused by the need to detect image near-duplicate in computer vision systems, as well as when image searching on Internet or in large digital archives. The main aim of the study: usage of multi-step segmentation for near-duplicate image recognition. The methods used in the study: clustering of image pixels brightness is used for segmentation. The recurrent neural network is used for clustering. To estimate images similarity the authors have applied the cosine distance between rank distributions of clusters cardinality. The results: The authors suggested the search patterns based on the rank distributions of brightness clusters cardinality. The paper introduces the experimental results on the near-duplicate image recognition based on application of the suggested search patterns. It is shown that the use of a multi-step segmentation and rank distribution of the brightness clusters cardinality allows determining reliably the near-duplicate of the original image with a high degree of distortion on them, up to the radius of the Gaussian distortion equal 8 pixels. Such an approach also allows solving reliably the inverse problem of detecting the original image even in its fivefold reduced copy with radius Gaussian distortion on it to 8 pixels.

Keywords

• image
• pixel
• point mapping
• recurrent neural network
• clustering
• segmentation