The main idea of the segmentation-based target proposal method is to combine the segmented image regions to obtain the target region. There is a problem: the error in the previous step will lead to the error in the next step, and this method of merging divided regions in one direction is flawed. Therefore, this paper adopts multi-branch method to expand the search scope and improve the accuracy.
It is not advisable to simply use greedy strategy to get the target area by searching the divided regional space. In this paper, a principle is adopted for the merging of segmented regions: complex targets with multiple color and texture regions need to use different merging strategies to complete the merging of segmented regions. The main idea of this paper is to improve the accuracy of selective search methods by combining various combination strategies.
The research content of this paper mainly consists of two parts: multi-segmentation region combination strategy and segmentation region search.
(1) learning complementary merger strategy
In this paper, the process of region merging is abstracted as a linear classifier, and the classifier is trained by changing the weight of each training sample (similar to the Boosting process). However, unlike Boosting, the region combination classifier proposed in this paper does not combine multiple weak classifiers into a strong classifier, but forms a new classifier through error correction between classifiers.
(2) Multi-stage branching
Through the above-mentioned region combination classifier, the image segmentation region can be searched, and a greedy aggregation step can be divided into multiple steps.
In this paper, the combined search space of segmented regions is increased by tree organization, and the SVM classifier is trained by the results of region combination.
In this paper, different regional combination strategies are organized by tree branches, and the division method of regional combination strategies under the same branch is the same.
Assuming that the height of the region aggregation tree is t and the degree of each non-leaf node of the tree is k, KT region aggregation results can be obtained at the bottom, totaling * * *, and a proposal window can be obtained, where λ represents the number of scattered regions added in each region aggregation, and n represents the number of segmented regions owned by the initial image. You can use λ and n to control the final number of suggestions.
The generation process of each branch is actually a greedy aggregation method, but this process is hierarchical.
The previous point describes how to branch, and then how to use the image segmentation region aggregation tree to generate suggestions, mainly how to design the segmentation region classifier.
In order to make the proposed classification model the same in different branches of the image segmentation region aggregation tree, a binary linear classifier is used to classify the segmentation regions.
That is, different regions belonging to the same object are taken as positive examples, and different regions belonging to different objects are taken as counterexamples as training. Then the corresponding SVM classifier is obtained.
The original image data of the experiment comes from PASCAL VOC2007. The initial image segmentation algorithm and image region features are exactly the same as selective search. This paper mainly improves the region aggregation algorithm.
The algorithms used for comparison are selective search (SS), geographic object suggestion (GOP), global and local search (GLS), edge box (EB) and binary normalized gradient (Bing).
The experimental results are compared as follows:
In contrast, our method is not much different from selective search (SS). When the number of suggestions is low, the results of edge box and selective search are better than the methods proposed in this paper. In the case of more suggestions, the method proposed in this paper has a slightly better effect.
The above table is a comparison of the recognition results of several methods in PASCAL VOC2007 plus R-CNN, and the IoU parameters in the comparison cases are not indicated in the article. As can be seen from the table, the algorithm proposed in this paper is not much different from selective search in retrieval accuracy, and the improvement of the algorithm proposed in this paper is very limited.
The main idea of this paper is to improve the selective search method in the process of segmentation region aggregation. In this paper, the previous single greedy aggregation process is divided into multiple greedy aggregation processes by using multi-layer tree aggregation method, and the regions are classified by SVM. The tree region aggregation method adopted in this paper does increase the search space of segmentation region aggregation, but it is still greedy aggregation from the point of view of a single process, so the article does not improve the selective search much, and mainly provides a new idea of segmentation region aggregation.
Compared with the experimental results, the method proposed in this paper is slightly better than the selective search only when the number of suggestions is small, but its spatial complexity is much greater than that of the selective search, and the time for calculating suggestions is increased.
Selective search for object recognition. International Journal of Computer Vision104.2 (2013):154-171.