Research on energy consumption optimization of RetinaNet model based on lightweight improvement in edge intelligent terminals in power system

With the increasing application of edge computing in power systems, intelligent terminal devices are facing new challenges in processing capabilities and energy optimization. As a highly efficient target detection model, RetinaNet's potential application in edge intelligent terminals of power systems has not been fully explored. This study proposes a lightweight model based on the convolutional neural network RetinaNet. Multiple lightweight models are used to replace the original backbone network ResNet for comparison. The best model is selected while ensuring accuracy. Redundant connections in the model network are pruned through channel pruning to reduce model size and improve detection speed. The results show that compared with the original algorithm, the RetinaNet model proposed in this study reduces parameter count by 73%, decreases computational load by 41.8%, reduces model size by 72.7%, and only decreases the mean average precision (mAP) value by 1.8 percentage points.