A Salient Region Detection Method Based on Joint Sparse Multi-Scale Fusion

Abstract

The invention belongs to the technical field of image salient area detection, and specifically discloses a method for detecting image salient area based on joint sparse multi-scale fusion, the steps of which include: (1) constructing a multi-layer Gaussian pyramid for training image sets to realize multi-scale, Training to obtain dictionaries at various scales; (2) Take an image block for each pixel in the test image, and jointly sparsely solve the sparse representation coefficient of the image block at each scale; (3) Use the sparse representation coefficient as a feature, and perform Calculation of saliency; (4) Fusing the saliency results at multiple scales to obtain the final saliency map. The present invention achieves the purpose of extracting the region of interest of the human eye in any given image, and its advantages are: firstly, the multi-scale operation overcomes the influence of different scales of the image; secondly, the joint sparse framework is very beneficial to the subsequent saliency calculation. Experiments show that the results of this method are robust and outperform the results of most existing methods.

Description

Technical field [0001] The present invention belongs to the technical field of image salient area detection, and can be used to extract the area of ​​interest of the human eye in any given image, and provide better work for subsequent video image compression, image segmentation, target recognition, image repair, image retrieval, etc. The reference information is specifically a method of image salient area detection based on joint sparse multi-scale fusion. Background technique [0002] 80% of the information that humans obtain from the external environment comes from the visual system. When facing a complex scene, human eyes will quickly turn their eyes to the areas of interest and give priority to these areas. Further processing, this special processing mechanism of the human eye is called the visual attention mechanism. In daily life, the human eye obtains a large amount of information every day, and processes and processes it automatically and efficiently. The visual attentio...

AI Technical Summary

Problems solved by technology

[0004] Although the current bottom-up salient region detection algorithm has achieved good results, most methods, such as the classic Itti method and SR method, have a very serious problem and defect in their calculation process

When detecting, it is easy to make the detection results concentrate on the edge of the target, and it is difficult to detect the entire salient target, because most bottom-up methods use the center-peripheral difference operation, because the pixels on the edge and the surrounding pixels The feature difference of the point is large, and the attention degree is high; while the pixel point located in the center area of ​​the target has a small difference with the surrounding pixel point feature, so the attention degree is low

Images