Improved YOLOv8 for oil manhole covers detection using attention mechanism and depthwise convolution

This paper presents an improved YOLOv8 detection framework that fuses Transformer and depth-separable convolution for the oil manhole covers detection task in the complex operating environment of offshore oil platforms, which suffers from insufficient feature characterisation capability and difficulty in identifying multiple small targets. Through an architectural optimisation strategy, the Transformer-based C3TR and Swin Transformer-based C3STR modules are introduced to enhance the multi-scale global context modelling capability by constructing cross-layer attention mechanisms. Secondly, the Depthwise Convolution (DWConv) is used instead of the standard convolutional layer to reconstruct the neck network to achieve a balanced optimisation of computational efficiency and recognition accuracy. Experiments on the oil dataset show that compared with the standard YOLOv8 model, the model has improved [email protected] by 0.8%, precision by 2.7%, and recall by 2.2% on the validation set. The ablation experiment further quantifies the contribution of each improved module. In order to verify the generalization ability of the model, this paper also conducted supplementary experiments on the public manhole cover dataset. The validation set indicator [email protected] has increased by 0.7% and precision has increased by 4.3%. The experimental results show that by coordinating the optimization of the attention mechanism and lightweight design, the generalization ability of the industrial inspection model under limited data conditions can be effectively improved, providing a valuable solution for promoting the construction of smart cities.