Comparative Evaluation of Deep Object Detection Architectures for Automatic Target Recognition in Real-World...

Reliable detection of underwater objects is important for mine countermeasures, unexploded-ordnance surveys, navigational-hazard assessment, and subsea infrastructure inspection. Interpretation of side-scan sonar (SSS) imagery is still largely performed by trained human operators, making post-mission analysis time-consuming when large volumes of data are collected. Automatic Target Recognition (ATR) can support this process by automatically detecting potentially relevant contacts and reducing the amount of imagery requiring detailed manual inspection. However, the ability of deep-learning detectors to generalize across real underwater missions remains insufficiently investigated. This study presents a comparative evaluation of five deep object-detector configurations, Faster R-CNN, RetinaNet, SSD, YOLOX, and DETR, for ATR using real-world SSS imagery acquired by autonomous underwater vehicles (AUVs). The development dataset comprised 12,066 image patches containing 18,253 annotated instances from three classes, acquired during five measurement missions. Two additional complete missions were excluded from all stages of model development and used as independent cross-mission tests of the same physical reference objects under previously unseen acoustic and acquisition conditions. Within-development validation ranked Faster R-CNN first, with a three-class mAP50:95 of 0.451 and a recalculated two-class value of 0.568; this validation ranking was used for model selection and was not treated as evidence of independent generalization. On the held-out missions, SSD achieved the highest primary mAP50:95 on Test A (0.614), whereas YOLOX achieved the highest value on Test B (0.084) and the highest inference throughput. A retrospective comparison of the validation-selected and terminal training checkpoints showed greater checkpoint sensitivity on Test A than on Test B, although SSD and YOLOX remained the highest-performing configurations on their respective missions. The substantial variation in detector ordering and performance between the held-out missions highlights the influence of mission-specific conditions on ATR performance and demonstrates the importance of evaluating sonar detectors on complete real-world missions rather than relying on within-development validation rankings as evidence of generalization.