|
In computer vision tasks, various types of objects exhibit distinct characteristics in images. By learning and synthesizing the commonalities present in the training set, the neural network effectively performs tasks associated with diverse objects. However, when the training set is incomplete—specifically, when certain classes are missing—it becomes challenging for the network to learn the features of these absent classes during testing. Consequently, the coverage of the training data must be evaluated prior to training the network. This study uses synthetic aperture radar (SAR) aircraft detection as an example to illustrate the importance of evaluating dataset coverage, introduce evaluation methods, and propose solutions for incomplete dataset. Variations in SAR target features occur when the relative observation angle of SAR changes, causing changes in the brightness of the target scattering points. SAR images can exhibit significantly different characteristics even for similar targets. Based on this characteristic, the aircraft in SAR images are classified into eight angle-based classes. If the training set includes fewer than eight angle types for aircraft (at least one), the network will be unable to detect aircraft from all eight angles in the test set. To tackle potential issues arising from incomplete training sets, the following solutions are proposed: Firstly, a clustering algorithm is employed to classify the labeled data more accurately by considering the differences in the heat maps of various feature data. Next, the average heat map is extracted for each data class, overlaid, and compared with the average heat map of the complete test set to identify any missing data types. Finally, the training set is supplemented with the appropriate data based on the identified missing data types. Experimental results using partial data from the SAR-AIRcraft-1.0 dataset demonstrate the effectiveness of the proposed method.
|
