Multiple Target Recognition Within SAR Scene Achieved Using YOLO and Explanability Investigated Using Gradient-Free Visualisation

Detecting multiple targets of interest within a Synthetic Aperture Radar (SAR) scene is the most commonly addressed problem in Automatic Target Recognition (ATR). As there are not many SAR scenes with multiple ground targets against a simple background, SAR scenes are first simulated using the SARBake algorithm. Compared to other object detection methods, the most popular method You Only Look Once (YOLO), which is known for its speed and accuracy, is used in this paper. The YOLOv3 network with transfer learning is investigated as a reference, followed by different versions of YOLOv5 (s,m,l) and the YOLOv7. The networks are compared to find the one with the best performance and minimum computation time. The YOLOV5s with the fewest trainable parameters compared to the other models, performs well with multiple ground target detections in terms of Mean Average Precision (mAP) and Recall. In addition, all targets from 10 classes at different locations within the test SAR scene are detected and classified with an Intersection over Union (IoU) and a confidence score always greater than 0.80. Later, the object identification and localisation predicted by the YOLOv5s on unseen test scenes are explained using a gradient-free visualisation technique called Score-CAM. A comparison is shown between the gradient-based and gradient-free visualisation methods to shed more light on how Score-CAM outperforms other explanations. This helps to translate different computations performed by Deep Neural Networks (DNNs) as a logical step that is more comprehendable to human intuition, such as spatial regions.