Abstract: In order to improve the detection accuracy and speed of convolutional networks, and reduce the model under miscellaneous scenes, a visual detection method of press connection pipe under complex field background based on wavelet analysis and improved lightweight neural network is proposed. The image noise is removed using wavelet analysis and the lightweight GhostNet model is used in the backbone network. In addition, omni-dimensional dynamic convolution is introduced to enhance the backbone feature extraction capability, depth separable convolution is used to reduce the model complexity, improved convolutional block attention module(CBAM)is embedded to improve the model accuracy, and K-Means ++ algorithm is used to cluster the anchor frame size and linear transformation to accelerate the convergence of the target frame. CIoU-NMS is used to improve detection speed and accuracy. The actual test results show that compared with YOLOv4 model, the improved YOLOv4 lightweight model size is significantly reduced by 199.7 MByte with the accuracy is only lost by 2.98%, the detection speed is increased by 3.4 Hz to 33.9 Hz and the edge deployment efficiency index is better. Therefore the improved lightweight network model achieves the best balance among detection accuracy, model size and detection speed. Finally, the detection results in the field with complex background and multiple scenes also show that the algorithm can well meet the detection requirements in practical engineering tasks, and has sound engineering practicability.

Key words: pressure connection pipe, complex background, improved convolutional network, lightweight, omni-dimensional dynamic convolution( ODConv)