Analysis of MSTAR Object Classification Features Extracted by a Deep Convolutional Neural Network

Introduction. Deep convolutional neural networks are effective tools for classifying objects on radar images; however, their decision-making process is not transparent. This makes the determination of classification features of objects that affect the entire network operation a relevant research task. Such a study contributes to the field of explainable AI (XAI). Aim. Determination of the classification features of military objects, detected by a deep convolutional neural network during training. Materials and methods. The convolutional neural network was designed, trained, and tested using Keras and Tensorflow 2.0 libraries on the open part of the MSTAR dataset. The GradCAM method was used to visualize and determine the classification features of objects in the dataset. Results. When using MSTAR images with an unsuppressed background, the object itself makes a significant contribution to the classification result only for 58 % of the images. For 6 % of the images, the classification result is determined by the object radar shadow, and for 25 % of images – by the background. For 11 % of the images, the most significant classification feature could not be established. For images with a suppressed background, in 60 and 40 % of the cases, brightness distribution and the object contour, respectively, made the main contribution to the classification result. Conclusion. In the MSTAR dataset, each class of objects is represented by a set of radar images of the same real object from different view angles. This determines local background features, invisible to humans and unique to each class of objects. These features have a significant effect of the training outcome of the neural network. This effect can be eliminated by suppressing the background and reducing the image dimensionality. The obtained results also suggest the feasibility of further research into the XAI capabilities in relation to modern neural detectors and radar image datasets.