Research on lightweight terminal mark detection method based on improved DBNet network

To address the intricate backdrop and distorted deformation issues in substation terminal marking identification, a lightweight detection method utilizing an enhanced DBNet network is proposed. To address the intricate background factors in the terminal marking image, the backbone network is substituted with the lightweight MobileViTv3, and the DCA module of the dual cross-attention mechanism is incorporated to capture both local details and global contextual information. The Dynamic Snake Convolution (DSConv) is implemented within the feature pyramid to dynamically modify the sampling paths of the convolution kernel, while the offset generation network is revised to an MLP for enhanced accuracy in offset generation. Additionally, the upsampling operation of the FPN layer is replaced with the lightweight upsampling operator CARAFE, which adjusts the upsampling kernel based on the input feature map content. Furthermore, the Dice loss function is integrated into the DBNet architecture to enhance network performance. The experimental findings indicate that the detection accuracy F1 of the enhanced lightweight DBNet network attains 93.4%, surpassing the original network by 4.9 percentage points, while the number of parameters is merely 23.6% of that in the original model, thereby adequately fulfilling the practical requirements for detecting twisted and deformed terminal markings in a complex background.