Oil Tank Detection With Improved EfficientDet Model
Su Xu, Haowei Zhang, Xiping He, Xiaoli Cao, Jian Hu
Abstract
Rapid detection of oil tank targets has become a topic of significant and increasing interest because the quick acquisition of the distribution and volume of oil storage tanks has high economic and military value. However, research in the literature still faces many challenges in different scenarios. This paper presents a new oil tank detection approach for improved EfficientDet model. The proposed approach has three primary operations: First, adding three-dimensional deformable convolution to the EfficientDet model as pre-detection to limit detection in the smaller area. Second, loading an attention mechanism appropriate for Oil Tank detection on the myNet model highlights tiny target information. Lastly, training the myNet model repeatedly by focal loss function to seek better results. The result shows that the overall accuracy of mean average precision in the proposed approach increased by at least 8.25% compared with that of the tested conventional approaches. Therefore, it provides advantageous capabilities for monitoring Oil Tanks in high-resolution remote sensing imagery.