Select the search algorithm

Relevant people who are engaged in the field of artificial intelligence and have target recognition research are not ignorant of the choice of search algorithms. Since the beginning of its introduction, the algorithm has achieved remarkable results in target recognition and positioning, and its impact has been more than ten years.

The following is the author’s 2012 paper “Selective Search for Object Recognition”, which was published in IJCV, to take you to understand the core idea of the algorithm and the value it embodies in target recognition.

Select the search algorithm

Figure 1 Study three questions

As shown in Figure 1, the authors begin by proposing a choice of search methods and give three questions about its research:

(1) What is the best choice search diversification strategy for the adaptability of segmentation?

(2) How effective is it to create high-quality locations for small data sets using the select search method in images?

(3) Can selective search bring stronger performance to target identification and model classification?

It can be seen that the research team not only considers whether the algorithm itself is feasible, but also pays more attention to the actual effect and performance of the algorithm. I believe that the small partners who have read this paper know why the author has this question.

At the same time, the author introduces the problems in image recognition at great length, and gives a multi-faceted comparison and evaluation of different solutions to the problem. Finally, the author would like to say: the image is essentially hierarchical, target recognition and positioning need to be separated, specifically can be carried out by combining segmentation and search technology, to complete the identification and positioning of the target, rather than through a single technology to achieve, should take different strategies to improve the target positioning accuracy, otherwise it will not work!

Select the search algorithm

Figure 2 Algorithm flow

As shown in Figure 2, we can clearly see the implementation process of the entire algorithm. What kind of idea is this pseudocode trying to convey? Just looking at the expression, do you find it too boring or difficult to understand? It’s actually very simple, let me give another vivid example. It is still a vegetable field, but this vegetable field has been carefully planned by Dabai, who planted different crops (targets) in different areas of the vegetable field to form a beautiful painting. The initialization process of the algorithm includes: the segmentation of the region of the image and the definition of the similar set S.

First of all, the regional segmentation method proposed by Felzenszwalb and Huttenlocher segments the image, which is equivalent to the great white standing on a high place overlooking the vegetable field, and at a glance, the vegetable field is composed of several areas composed of beautiful paintings, and each area is numbered in turn to get a collection.

Second, according to the order of numbering, the similarity size of each two adjacent regions is calculated, and the calculated similar result values are stored in the set S. It is equivalent to the need to fertilize the vegetable field, but he considers that different crops need to apply different fertilizers, and in order to be able to apply the correct fertilizer to the crops in the area, he compares the color, texture and size of the same crop in the adjacent area in sequence, if these properties are more similar, the more similar the two areas are, and vice versa, the more similar, and the calculation results are stored in the set S.

Next, find the most similar value from the set S, which represents the most similar region of two adjacent numbers, assuming that the two regions are the sum, then the two regions can be combined into a region numbered: .

At the same time, the similarity of the merged regions is calculated separately from the areas adjacent to the number, and the similar values are stored in the set R. It is equivalent to The Great White after comparison, found that there is a high probability of planting the same crop in two areas, and he needs to use the corresponding fertilizer to fertilize. But he also wondered whether the same crop was planted in the two adjacent areas of the left and right adjacent areas of the merged area. The great white simply compares the merged area with the crops in the adjacent areas for color, texture, and size calculations, and notes the new results into collection S. Also, delete the similar values calculated by the regions to their neighbors before they were merged. In order to be able to obtain a definite fertilization area, the great white repeats the comparison to find two adjacent areas with high similarity, merging them into one area until all the areas in the vegetable field are merged.

Finally, all the areas of the merge are obtained and the corresponding boundaries of each region are obtained as the definitive areas for a certain crop.

Select the search algorithm

Figure 3 Excerpts from papers

At this point, the story has been introduced, I believe that everyone should understand the implementation of the algorithm. As shown in Figure 3, the authors in this paper by using a variety of strategies and methods to compare, and determine that the combination of regional segmentation of the selection search algorithm for the generation of the target location is very effective, and the use of multiple strategies can be applied to the image under various conditions.

Reference papers: Uijlings J R R, Van De Sande K E A, Gevers T, et al. Selective search for object recognition[J]. International journal of computer vision, 2013, 104(2): 154-171.

Be the first to comment

Leave a Reply

Your email address will not be published.