Article Preview
Top1. Introduction
Region localization is one of the important research directions in the field of computer vision. The goal of region location is to determine the position of an object region in the remote sensing image. At present, the common localization method is to use the supervised learning algorithm (Bahrami et al., 2021; Liu et al., 2016; Teng et al., 2017) to complete the object region localization in the test set according to the region category and location information training. The task of object region localization is to find out all interested regions in the image and determine their position and size, which is one of the core problems in the field of machine vision. Region localization is used in many scenarios, such as airport, harbor detection. Due to various regions have different appearance, shape, posture, imaging light and shielding factors, region localization has been the most challenging problem in machine vision. In many practical applications, such as small object detection (Elakkiya et al., 2021; Gong et al., 2021), traffic object detection (Fan et al., 2021; Ye et al., 2020), multi-modal object detection (Yin et al., 2018), medical object detection (Xi et al., 2020) and other tasks, data shortage and numerous missing marks cannot meet the requirements of neural network detection tasks. And in these applications, the data set differences between categories are bigger, the part of the unlabeled objects seriously polluted the background feature space. The classifier is difficult to distinguish the differences between the known categories and the current object. It makes an error classification, which confuses the judgment ability of the supervision model and leads to the lower accuracy of the model. As shown in figure 1, traditional region detection methods are divided into three steps. 1) Region selection (Tang et al., 2000; Yin & Li, 2020). That is, it uses sliding window to cover a part of the image to be tested as a candidate region; 2) Feature extraction (Guan et al., 2021; Karim et al., 2020; Li et al., 2022). The features are related to the candidate region, such as HOG, SIFT; 3) Classification. It uses the classifier completed by training to classify classes, such as the commonly used Support Vector Machine (SVM) model, Adaboost, DPMC6, RF(Rodriguez-Galiano et al., 2012) (random forest).
Figure 1. Flowchart of traditional region localization
However, these algorithms all need to manually obtain the relevant region feature information from the original image with many limitations (Daura-Oller et al., 2009; Yin & Li, 2021; Zhou et al., 2002):
- 1.
Poor portability. For specific detection tasks, different methods need to be designed manually. For different regions or different forms of the same region, designers have high requirements for experience.
- 2.
The classification of feature extraction and training is a common problem of traditional detection models. If the extraction of artificial features occurs omission phenomenon in the design process, the missing useful information will not be recovered from classification training, thus affecting the detection results.
- 3.
The traditional methods mostly adopt sliding window to conduct traversal search, and divide the image into small blocks of various sizes, and then identify the image blocks. It retains the part with high probability, and merges or deletes the part with low probability. The complexity of this method is high. There are a large number of redundant small pieces, which seriously affect the operation speed, and it is difficult to achieve in reality. Therefore, since the rise of deep learning in the field of target detection in 2013, it has quickly replaced the status of traditional algorithm.