SOD-HRFF: Small object detection algorithm based on high-resolution
image processing and fusion of different scale features
Abstract
To address the issue of misidentifying and omitting small targets when
using UAV high-resolution aerial photos for small-target detection
tasks, as a result of the limited percentage of small target pixels and
vulnerability to background noise interference. This paper introduces a
small target detection algorithm based on high-resolution image
processing and fusion of different scale features. ObjectBox is adopted
as the baseline network. Firstly, the high-resolution image processing
module (HRIP) is introduced to extract small targets’ spatial and edge
features. Second, the LeakyRelu activation function is used in the
ordinary convolution, so that the network can maintain a specific
response in the negative range and maintain the gradient in the small
range of eigenvalues. Finally, the Bidirectional Feature Pyramid Network
(BIFPN) is used to realize the multi-branch different scale feature
fusion to alleviate the mutual occlusion problem due to the dense
distribution of the small targets, and to improve the model’s ability to
locate the bounding box of the small targets accurately. Experiments on
the VisDrone2019 dataset prove that by enhancing the baseline model, the
small target detection accuracy is improved by 0.145, proving the
proposed algorithm’s effectiveness.