Remote Sensing Imagery Object Detection Model Compression via Tucker Decomposition

• DOI: 10.3390/math11040856
• Publication Date: 2023-02-08T09:57:41Z
• Authors: Lang Huyan; Ying Li; Dongmei Jiang; Yanning Zhang; Quan Zhou; Bo Li; Jiayuan Wei; Juanni Liu; Yi Zhang; Peng Wang; Hai Fang
• Keywords: Initialization; Rank (graph theory); Representation; Speedup

Although convolutional neural networks (CNNs) have made significant progress, their deployment onboard is still challenging because of complexity and high processing cost. Tensors provide a natural compact representation CNN weights via suitable low-rank approximations. A novel decomposed module called DecomResnet based on Tucker decomposition was proposed to deploy object detection model satellite. We remote sensing image compression framework which consisted four steps, namely (1) initialization, (2) initial training, (3) the trained reconstruction model, (4) fine-tuning. To validate performance in our real mission, we constructed dataset containing only two classes objects DOTA HRSC2016. The method comprehensively evaluated NWPU VHR-10 CAST-RS2 created this work. experimental results demonstrated that method, Resnet-50, could achieve up 4.44 times ratio 5.71 speedup with merely 1.9% decrease mAP (mean average precision) 5.3% dataset.