Identification of Activated Sludge Microbial Based on Improving YOLOv8

Abstract

Manual microscopic inspection of activated sludge is costly, expertise-intensive, and prone to errors, limiting its widespread use. To address these challenges, we propose an improved visual model based on YOLOv8n. This model integrates an optimized convolution module in the backbone network, incorporating sensory field convolution and Coordinate Attention (CA) mechanisms to enhance feature extraction and detection in complex backgrounds. The C2f module is refined with convolutional reparameterization (OREPA), reducing computational complexity while maintaining accuracy. The MPDIoU function is augmented with the facoler-IoU loss function, improving delineation of challenging samples and enhancing learning capability.Testing on an activated sludge microbial image dataset shows our model achieves 90.3% and 95.7% accuracy, an increase of 5.9% and 4.7%, respectively. After overall optimization, the model achieved an mAP of 90.3%, which is 6.9% higher than YOLOv8n, with an FPS of 105.3, representing a 24.6% improvement over YOLOv8n. Additionally, the average inference time is reduced by 1.3 ms. When evaluated on Kaggle’s algal cell dataset, recognition accuracy improves by 4.9%. These results demonstrate the model’s effectiveness in reducing omission and misdetection in activated sludge microscopy. This work provided new ideas and methods in the field of activated sludge microorganisms.

Keywords

• Activated sludge detection
• YOLOv8n
• feceptive field attention
• convolution parameterization