A multibeam water column image (WCI) can provide detailed seabed information and is an important means of underwater target detection. However, gas plume targets in an image have no obvious contour information and are susceptible to the influence of underwater environments, equipment noises, and other factors, resulting in varied shapes and sizes. Compared with traditional detection methods, this paper proposes an improved YOLOv7 (You Only Look Once vision 7) network structure for detecting gas plume targets in a WCI. Firstly, Fused-MBConv is used to replace all convolutional blocks in the ELAN (Efficient Layer Aggregation Networks) module to form the ELAN-F (ELAN based on the Fused-MBConv block) module, which accelerates model convergence. Additionally, based on the ELAN-F module, MBConv is used to replace the 3 × 3 convolutional blocks to form the ELAN-M (ELAN based on the MBConv block) module, which reduces the number of model parameters. Both ELAN-F and ELAN-M modules are deep residual aggregation structures used to fuse multilevel features and enhance information expression. Furthermore, the ELAN-F1M3 (ELAN based on one Fused-MBConv block and three MBConv blocks) backbone network structure is designed to fully leverage the efficiency of the ELAN-F and ELAN-M modules. Finally, the SimAM attention block is added into the neck network to guide the network to pay more attention to the feature information related to the gas plume target at different scales and to improve model robustness. Experimental results show that this method can accurately detect gas plume targets in a complex WCI and has greatly improved performance compared to the baseline.

Artificial reefs serve as a crucial measure for preventing habitat degradation, enhancing primary productivity in marine areas, and restoring and increasing fishery resources, making them an essential component of marine ranching development. Accurate identification and detection of artificial reefs are vital for ecological conservation and fishery resource management. To achieve precise segmentation of artificial reefs in multibeam sonar images, this study proposes an improved YOLOv11-based model, YOLO-AR. Specifically, the DCCA (Dynamic Convolution Coordinate Attention) module is introduced into the backbone network to reduce the model’s sensitivity to complex seafloor environments. Additionally, a small-object detection layer is added to the neck network, along with the ultra-lightweight dynamic upsampling operator DySample (Dynamic Sampling), which enhances the model’s ability to segment small artificial reefs. Furthermore, some standard convolution layers in the backbone are replaced with ADown (Advanced Downsampling) to reduce the model’s complexity. Experimental results demonstrate that YOLO-AR achieves an mAP@0.5 of 0.912, an intersection-over-union (IOU) of 0.832, and an F1 score of 0.908. Meanwhile, the parameters and model size of YOLO-AR are 2.67 million and 5.58 MB. Compared to other advanced segmentation models, YOLO-AR maintains a more lightweight structure while delivering a superior segmentation performance. In real-world multibeam sonar images, YOLO-AR can accurately segment artificial reefs, making it highly effective for practical applications.

Facial occlusion removes expression-relevant evidence and remains a major source of error in camera-based affective sensing. Existing approaches to occlusion-robust facial expression recognition often rely on specialized attention, reconstruction, semantic, or geometric pipelines, whereas aggressive training using synthetic occlusions may impair discrimination on clean images or overfit to synthetic corruption patterns. We address this tension between cleanness and robustness through clean-anchored hard occlusion fine-tuning (CA-HOFT). HardMix samples structured and random occlusion modes according to a facial region-weighted distribution. An explicit classification branch for the non-HardMix source view preserves ground-truth supervision, while a fixed reference teacher initialized from the preceding mixed-occlusion stage supplies a stationary distribution for both paired student views. These signals jointly train a single classifier while retaining a single-backbone inference pathway. Across five independent training runs, the ResNet-18 student achieves 90.08% accuracy on the Real-World Affective Faces Database (RAF-DB) and 87.38% accuracy with 83.34% macro-F1 on Occlusion-RAF-DB, with corresponding sample standard deviations of 0.21, 0.12, and 0.32 percentage points. The retained model has 11.18 million parameters and requires 1.814 giga multiply–accumulate operations (GMACs). Controlled comparisons and ablations indicate that it has the most favorable observed cleanness–robustness trade-off among the tested epoch-matched alternatives; however, fixed-checkpoint comparisons on Occlusion-RAF-DB are not significant after Holm correction. AffectNet-8 and evaluations using natural occlusion provide supporting evidence, while broader cross-domain validation remains future work.