Year: 2024
Type: article
Abstract: An improved YOLOv8 model is proposed to address the issue of poor recognition performance caused by their low resolution and weak feature representation in small object detection task. Firstly, to extract a richer set of low-level features from images, a WT_Conv module is designed to fuse the feature components extracted by WT (Wavelet Transform) with those extracted by convolutional layer. Secondly, based on the idea that shallow and deep feature maps contain information at different scales, a MS (Multiscale Separation) module is designed to preserve the features of small objects separated from shallow layer and transfer the salient features of large objects to the deeper layers, effectively solving the problem of inconsistent feature expression caused by the direct fusion of shallow and deep feature maps. Finally, we introduce the DE (Detail Enhancement) module capable of fusing adjacent feature maps to process the small-objects features separated by the MS module, enhancing feature representation for small objects. Experiment results on UAVOD-10 and Small Object datasets show that our model achieves a mAP improvement of 9.5% and 2.3% respectively over the baseline, and it also shows a significant advantage over other comparative models, affirming the effectiveness of the proposed model for small object detection tasks.
DOI: 10.54097/n5rtnt71
Authors: Yingying Tan, Jinpeng Song, Chen Chu
Institution: North China Electric Power University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4405901948