Sli-EfficientDet: A slimming and efficient water surface object detection model

Abstract

In the field of water surface object detection, deep learning technology has become a mainstream method. Unmanned Surface Vehicles (USVs), which perform precise sensing and measurement tasks on water surfaces, particularly benefit from these advancements. However, for hardware resource-constrained USVs, current detection models still struggle to find a balance between being lightweight and maintaining accuracy. To address this challenge, we first reduce parameters by clipping channels in the backbone network through a dependency graph based pruning method. Additionally, we introduce the Simple Attention Module (SimAM) into the backbone network to derive excellent three-dimensional attention weights without adding additional parameters during computation. Furthermore, we utilize the ghost module to reconstruct the feature fusion network by using simple linear operations to process feature maps, which enhances the network performance in feature extraction while further compressing the model. Experiments show that our model achieves a 15.56 % improvement in mean Average Precision (mAP) while reducing the count of model parameters by 55 % compared to the original EfficientDet-D0 model, and balancing lightweight and accuracy compared to the majority of current models.