Selective search for object recognition paper notes [image object segmentation]

Reference resources  Selective Search for Object Recognition Paper notes 【 Image object segmentation 】 - cloud + Community - Tencent cloud

Preface ：

This paper introduces , If you quickly find the area that may be the object target , It is not like using traditional sliding windows to identify regions . Here, we use an algorithm to find images from multiple dimensional pairs , Possible regional targets , Reduce target fragmentation , Improve the efficiency of object detection .

Introduction and introduction ：

The picture is hierarchical , As shown in the picture below a：

　　　                                    　　　

Salad and spoon in salad bowl , And the bowl is on the table , In addition, the table is related to wood or everything on the table , So the goals in the picture are hierarchical . Image segmentation should be done hierarchically , There is no such general method as using a single strategy to segment images , Therefore, image segmentation is based on multiple strategies , But this will cause conflicts when merging regions . For example, in the picture above b chart , Cats can use color for segmentation , But their texture is the same . Contrary ; chart Ｃ The chameleon in is similar in color to the surrounding leaves , But the texture is different . Last , chart d in , Car wheels and cars are different in color and texture , But it matches the shape of the car very well . For these three graphs , It is impossible to segment images using one of their visual features .

In this article , The author combines intuitive segmentation algorithm and exhaustive search algorithm to propose this selective search( Selective search ) Algorithm . The intuitive segmentation algorithm is used to achieve the segmentation from bottom to top in combination with the structural level of the picture , To generate the target area . The purpose of using the exhaustive search algorithm is to get all the possible target areas . Selective search algorithm , Using a diversified resampling algorithm .

In this article , The author mainly introduces the selective strategy from the following questions ：

1.   What kind of diversity strategy does the selective strategy adopt to adapt to the segmentation of images ？

2.   How efficient is the selective strategy in generating high-quality small targets in images ？

3.   Can we use selective strategy to combine classification model and appearance model for target recognition ？

Introduction to selective algorithm ：

Feature introduction ：

1. Applicable to all sizes .

The target can appear in the picture in any size , Even the boundary between some goals and others is not obvious , In the face of these problems , The selective algorithm will record all target sizes , Just like the picture below , It can be easily implemented using hierarchical algorithms .

                                

2. diversification .

A single strategy cannot deal with all kinds of differentiation areas . So a variety of strategies are used, such as color space , texture , Coincidence degree, etc .

3. Fast calculation .

Process introduction ：

  The selective algorithm uses a hierarchical merge algorithm （Hierarchical Grouping）, The basic idea is ：

By dividing a picture from low to high , When we draw a large area , Continue to iterate over this large area , Until no area can be drawn . And record all the areas produced in this process , Through the color , texture , Coincidence , Size to merge these fragmented areas . There is no need to set the sliding window in this way , Sliding grid , Can be adapted to any target size .

So the specific process of this algorithm ：

　                                             　　

1. use first Efficient Graph-Based Image Segmentation The method in this paper is to quickly get the segmentation region according to the hierarchy Ｒ;

2.  Initialize similarity set Ｓ;

3. From the set of partitioned regions Ｒ Calculate the similarity in pairs , Put it into the similarity Ｓ Collection ;

 4. From the similarity Ｓ Collection , Take out the two segmentation regions with the highest similarity . Then merge the two areas , And into the Ｒ in , Then from the similarity Ｓ Remove... From the set   Remove the area associated with the two divided areas . Then calculate the merged New Area   And its adjacent areas , Put in Ｓ in , This cycle . until Ｓ Set is empty ;

５.  repeat ３ Until this area becomes a ; Then output all the changed areas in the process ;

About diversity strategies ：

It is divided into two parts ： Color space diversity 、 Regional similarity diversity .　

1.   There are eight kinds of color space diversity ：

[1]RGB

[2]I grayscale （grey) 

[3]Lab

[4]RGB Normalized in the image rg Grayscale images of channels and images

[5]HSV

[6] The normalized rbg

[7]C

[8]H

2. Regional similarity diversity ：　

To texture 、 Coincidence 、 The size of these features are calculated , Specific push down process , See paper , So how to use selective algorithm in object recognition ？

　　　

We use a selective algorithm to obtain a series of possible target areas L, Then we will label the target area in advance （ We become GT） As a positive sample , stay Ｌ In the set GT Of IOU stay 0.2~0.5 As a negative sample of this class , For coincidence degree and IOU exceed 0.7 The negative sample of , I throw it away , Then the data of these areas , Feature extraction , Used in the paper SIFT Algorithm , Then these features are put together into linear SVM Do this type of training . And then the samples with high scores , Put it into the negative sample to continue training . In turn .

Article address ：خطای 404 : تی پی بین

Reprinted address ：Selective Search for Object Recognition Paper notes 【 Image object segmentation 】 - Gong Xijun - Blog Garden