Sewage and Location Detection with Improved Cycle Generative Adversarial Network-Based Augmented Datasets and YOLOv5-BiFC

Abstract

Sewage discharge from outfalls significantly contaminates the environment. However, due to the unique characteristics of environmental policy, challenges such as data acquisition difficulties arise. This study introduces an enhanced approach by utilizing an improved Cycle GAN, the core function of which involves extrapolating a small sample to a large sample. An enhanced YOLOv5 model is used to focus on lightweight model construction and model performance enhancement. The proposed Cycle GAN incorporating Self-Attention and residual modules is suggested to tackle the problem of limited data at the outfall. Additionally, a refined YOLOv5 model (YOLOV5-BIFC) is proposed, integrating the C3Ghost module, BiFPN module, and CBAM attention mechanism to overcome low model recognition efficiency and large model size concerns. The research employs an augmented dataset for training and evaluates model performance using metrics such as mAP, F1 score, model size, and FPS. The results indicate that the YOLOV5-BIFC model has a size of 7.8 MB, representing a 44.2% reduction compared to the original YOLOv5 model. The FPS of this model is 26.7, enabling real-time discharge monitoring. The F1 score and mAP achieved by YOLOv5-BiFC are 89.8% and 87.3%. Comparative analysis with other neural networks demonstrates the superior accuracy and efficiency of the enhanced model. The YOLOv5-BiFC model facilitates precise, rapid, and intelligent discharge inspection.

Keywords

• sewage detection
• deep learning
• target detection
• YOLOv5
• generative adversarial network