Litcius/Paper detail

Underwater debris detection using YOLOv12 with enhanced feature extraction using R-ELAN and FlashAttention network

N. Deluxni, Pradeep Sudhakaran

2025Results in Engineering7 citationsDOIOpen Access PDF

Abstract

• The YOLOv12 model with R-ELAN and FlashAttention enhances detection accuracy. • The SPPF module improves multi-scale feature fusion for better object recognition. • The model achieves 84.6% mAP@50, 83.9% Precision, and 81.7% Recall. • The 14 ms inference speed ensures real-time maritime waste monitoring efficiency. Detecting undersea debris is crucial to attempts to clean the ocean and preserve marine ecosystems. Conventional object detection techniques are difficult to implement underwater cleaning vehicles because of poor visibility, variable lighting, and refraction distortions in water. This research presents an enhanced technique for real-time underwater debris identification utilizing YOLOv12 model. The proposed model combines R-ELAN (Recurrent Efficient Layer Aggregation Networks) with the FlashAttention Mechanism and the Spatial Pyramid Pooling - Fast (SPPF) module to enhance multi-scale feature fusion for enhance object recognition accuracy and effective feature extraction in challenging underwater environments. The proposed model also guarantees elevated detection accuracy across diverse item dimensions of the debris. The training model was validated by ten-fold cross-validation and ablation study, ensuring consistent performance across multiple data sets. The assessment criteria give the results of 84.6% for Mean Average Precision (mAP@50), 0.72 for Intersection over Union (IoU), 83.9% for Precision, and 81.7% for Recall. The model's inference speed of 14 ms per image on GPU which is more appropriate for real-time maritime waste monitoring applications. The findings indicate that the revised YOLOv12 architecture enhances the detection precision while preserving processing efficiency, providing a scalable solution for underwater cleaning and surveillance vehicle.

Topics & Concepts

Computer scienceFeature extractionUnderwaterObject detectionArtificial intelligenceFeature (linguistics)Pyramid (geometry)Computer visionSegmentationInferenceScalabilityIntersection (aeronautics)PoolingObject (grammar)Pattern recognition (psychology)Image segmentationIdentification (biology)Sensor fusionData miningCognitive neuroscience of visual object recognitionRemote sensingReal-time computingVisibilitySampling (signal processing)Advanced Neural Network ApplicationsWater Quality Monitoring TechnologiesUnderwater Acoustics Research
Underwater debris detection using YOLOv12 with enhanced feature extraction using R-ELAN and FlashAttention network | Litcius