Lightweight model of remote sensing ship classification based on YOLOv7-tiny improvement

Abstract

A lightweight ship classification and detection approach based on modified YOLOv7-tiny is presented to address the issue that the classification and detection of ships in optical remote sensing images is prone to mistakes and missed detection due to the complicated background in the offshore. Due to the uneven distribution of data set samples, this method applies a lightweight feature extraction module E_DSC to the backbone network to reduce the number of parameters and calculations of the feature extraction network and improve network reliability. On the other hand, the MPDIoU loss function is introduced to improve the network prediction box coverage and improve the network prediction accuracy. On a dataset of remote sensing images of ships, this strategy has been tested. The findings reveal that it improves the network’s average accuracy of ship recognition and classification by 3.3%. The amount of parameters and calculation is also better than YOLOv7-tiny with a 3% reduction in parameters and a 22% reduction in the calculation.