Johnson first proposed the concept, theory, and method of suitable-matching areas’ selection in 1972 [21]. The purpose is to select multiple sub-areas with good adaptability and specific sizes in a large-size reference image as navigation reference images for scene matching. The suitable-matching areas generally satisfy the following four points:

• Richness: The richer the image information contained in the selected area, the more conducive it is to the later matching calculation;
• Stability: Differences in imaging quality will cause changes in scene features, so matching in areas with unstable features is more likely to fail;
• Uniqueness: If the selected area contains similar targets, it will increase the probability of mismatching;
• Salience: Areas with significant features usually have noticeable feature differences from surrounding areas, which is helpful for distinguishing the foreground from the background.

How to analyze the matching ability of the image area and determine the selection criteria has always been the focus of research on scene-matching navigation technology. There are three main methods for selecting the good areas: manual selection methods based on experience, hierarchical screening methods based on multiple feature indicators, and pattern classification methods based on machine learning.

Since the manually selected suitable-matching areas have yet to pass scientific verification and lack objectivity, many researchers hope to perform area screening by extracting appropriate image features so that the selected areas meet the requirement. The hierarchical screening method is the most mature technology for selecting suitable-matching areas. It sets different image features as feature indicators and then uses pattern classification or multi-attribute decision-making methods to establish the mapping between scene adaptability and feature indicators. In [22], it constructs a metric function by fusing edge density, average edge intensity, and edge direction dispersion to solve the problem of automatic suitable-matching areas’ selection in local textureless target tracking. A method based on information entropy is proposed in [23], which uses improved information entropy, Frieden gray entropy, and normalized average mutual information as feature indicators. To solve the problem of changes in matching probability caused by noise in real-time images, ref. [24] improves and proposes two new indicators: phase correlation length and effective contour density. It also offers a matching suitability analysis method based on the fusion of multi-indexes by evidence theory. In [25], the combined weighting method calculates the combined weight of multiple feature attributes and selects suitable SAR (Synthetic Aperture Radar) scene-matching areas based on the comprehensive evaluation value.

The hierarchical screening method is easy to implement and highly interpretable. However, in constructing the mapping function, only the impact of feature indicators on the matching probability is considered, and the connection between different features is ignored. In addition, the fixed feature indicators and thresholds will cause the algorithm to be unable to cope with different types of scenes, making the algorithm less robust.

In recent years, some work has been performed on suitable-matching areas’ selection models based on machine learning. In [26,27], researchers used different image features to establish mapping relationships and transformed the selection problem into a clustering discrimination problem. They built SVM classification models to distinguish scene-suitable-matching areas from non-suitable-matching areas automatically. The essence of these methods is still to use traditional image feature parameters to train the classifier to predict adaptability. The trained model can present stable selection results for a particular type of image, but it still needs better versatility. On this basis, a matching probability prediction model, based on the ResNet deep learning network, is designed in [2,28] to guide the selection of the suitable-matching areas by predicting the matching probability of the subgraph. To enhance the robustness of the model, ref. [29] subdivides image data according to typical scenes and selects the suitable-matching areas from the perspective of multi-classification problems. Methods based on deep learning make the model more robust and more flexible. However, these methods require a large amount of annotated images as support. The output results are also closely related to the performance of the matching algorithm used during training, and the overall interpretability is poor. Unbalanced sample categories are prone to occur when the amount of data is small. Therefore, how to use deep learning to conduct adaptability analysis with a small number of samples is a great test for the network’s generalization ability.