Year: 2024
Type: conference-paper
Abstract: Remote sensing images are characterized by complex backgrounds and small targets due to the shooting angle and imaging, which are not conducive to target detection. The current mainstream algorithms use convolution for feature extraction and feature fusion, but the convolutional local receptive field is limited and tends to ignore the global information, which is not conducive to accurately differentiate the target from the complex background. To address these issues, we propose two modules for feature fusion: the lightweight LightMLP, which is used to capture global information of features to assist feature pyramid fusion features; and the channel spatial context-aware module CSCAM, which utilizes global context information across channels and across space to assist in distinguishing targets from backgrounds. Yolov8 is chosen as the base framework for comparison experiments and ablation experiments of the improved algorithms by reconstructing the dataset, and the results demonstrate that our proposed modules have good performance in improving detection accuracy, precision and recall.
Authors: Yong Sun, Yi Wang, Jie Xu, Yuying Zhang, Mingming Yang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4409830058