Design and verification of an explainable target recognition architecture for water surface situation awareness

Objective Existing vision-based water surface target recognition systems often integrate image enhancement, target detection, and result interpretation as separate stages. Such implementations limit module interchangeability, may leave enhancement objectives misaligned with detection requirements, and provide limited means of examining the spatial evidence behind predictions. This study proposes a layered and decoupled explainable target recognition architecture to address these issues.

Methods The architecture comprises four functional layers: perception input, feature enhancement, intelligent recognition, and decision interpretation. Unified interfaces specify input–output data formats and semantics, allowing functional modules to be configured, replaced, and combined. At the recognition layer, reusable network components are organized hierarchically and instantiated from configuration descriptions. For low-visibility scenes, an AOD-Net dehazing module is coupled with YOLOv3 and jointly optimized using the detection loss. A lightweight pixel-wise gating mechanism adaptively combines the original degraded image and the enhanced image before detection, retaining information from both sources. To examine alternative implementations within the same architectural framework, a separately trained AOD-Net–YOLOv3 cascade and IA-YOLO, which uses differentiable image processing and a convolutional parameter predictor, are also instantiated. For explainable analysis, a detection-tracing procedure associates each final detection after non-maximum suppression with a corresponding network prediction. Grad-CAM is then used to visualize detection-specific spatial responses, with four-panel outputs showing the input, detections, heatmap overlay, and prediction provenance. These responses are quantitatively evaluated using Pointing Game and Energy-based Pointing Game (EBPG). Experiments are conducted on SeaShips7000 using depth-aware synthetic fog images at five visibility levels, with training restricted to clear and light-fog conditions.

Results On the complete test set, the jointly optimized model achieves an mAP@0.5 of 0.5893, exceeding the YOLOv3 baseline (0.5705) by 1.88 percentage points and the independently trained cascade (0.5647) by 2.46 percentage points. It also achieves the highest precision (0.8816) among the four configurations. IA-YOLO yields the highest overall mAP@0.5 (0.5923), recall (0.3105), and F1 score (0.4560). Under moderate and heavy fog, the jointly optimized model attains the highest mAP@0.5, reaching 0.6065 and 0.5672, respectively. Its inference rate is 68.0 frames/s on the experimental GPU. The joint model maintains EBPG values between 0.9153 and 0.9268 across visibility conditions, while IA-YOLO exhibits the highest Pointing Game accuracy at all five levels.

Conclusion The results demonstrate that the proposed architecture supports configurable integration of different enhancement–recognition implementations and enables spatial analysis of individual detection results. Task-driven joint optimization improves detection performance over independent enhancement, while the complementary interpretability metrics reveal differences in response localization and concentration across models.