LMACNet: lightweight multi-scale attention convolutional network for glass insulator defects

Glass insulator defect detection faces challenges such as noise data interference, small defect targets, and low accuracy. Current most methods aimed at enhancing the accuracy of insulator defect detection inevitably increase model complexity, requiring additional computational resources. To address this issue, this paper proposes a lightweight multi-scale attention convolutional network (LMACNet). A general Spatial- Channel Cross Attention (SCCA) is proposed to perform cross-spatial learning on multi-scale features. SCCA- Mobile Inverted Bottleneck Convolution (SCCA-MBConv) is designed for lightweight feature extraction. Additionally, we improved the Cross Stage Partial Bottleneck with 2 convolutions (C2f) module to enhance multi-scale feature fusion and high-level semantic feature learning. Specifically, an auxiliary network is designed to train the lightweight model, alleviating the information bottleneck and ensuring sufficient gradient flow. Extensive experiments were conducted on the public Vietnam Public Merged dataset of Broken Glass Insulator (VPMBGI) dataset and a custom dataset, the results demonstrate that LMACNet achieves detection performance comparable to other related detection models, while having a parameter count of only 2.54M and requiring 3.01G computations. LMACNet exhibits efficiency and practicality.