An Efficient End-to-End CNN Network for High-voltage Transmission Line Segmentation

Abstract

Automation detection of power transmission lines is of great importance for intelligent power inspection, which could well serve the route programming and motion guidance of examination platforms. However, due to complex factors, such as complex natural environment, illumination change, image noise, efficient detection of transmission lines still frontages great challenges. Lately, deep learning has exhibited a good detection effect among different segmentation tasks. Nevertheless, it still has a few disadvantages in high-precision image segmentation, like inadequate detection, information loss caused by multiple pooling operations, etc. To realize automatic and accurate pixel-level extraction, an attention fusion segmentation network is put forward to provide an end-to-end segmentation module. Considering the the problem of class imbalance, a global attention model is introduced to make the module focus more on the target region and suppress the unimportant features. Meanwhile, aimed at the semantic gap, residual path is also proposed to achieve effective usage of local information. In addition, to solve information loss issue which arise from plenty of pooling processing, an attention fusion block is put forward to realize effective feature aggregation of multi-scale features and improve the detection ability of segmentation network on multi-scale objects. Experiments exhibit that the attention fusion segmentation network has a good extraction capacity compared with other classical segmentation network.