An Improved DBNet Model with Pyramid Pooling and Multi‐Scale Enhancement for Small Text Object Detection
Year: 2024
Type: article
Abstract: Detecting small text objects has been a key focus in object detection research due to their unique characteristics: small size, limited semantic information, susceptibility to interference in complex scenes, and tendency to be easily obscured, among others.At present, there are still two common issues in representative object detection models: First, small objects are easily interfered by the background or other objects, and second, multi-layer feature networks cause the loss of small object feature information.To address these challenges, this paper proposes an improved version of the DBNet model by introducing two modules: the contextual information fusion module SPP-CIF and the multi-scale feature enhancement module DA-MSFE.SPP-CIF fuses global and contextual information, by replacing the pooling layer of a pyramid with two sequentially concatenated dilated convolutions of small expansion rates, to encode semantic information of high-level features at multiple scales.DA-MSFE employs spatial attention and channel self-attention to select critical features at different scales and locations, and mines and exploits the correlations between channels to enhance and dynamically aggregate multi-scale features.Extensive experiments were conducted on the publicly available datasets MSRA-TD500 and ICDAR2015.The experimental results show that compared to the baseline model, the proposed model exhibits significantly superior performance in terms of the evaluation metrics.
Authors: Zhu Yuan, Wuxuan Tang, Jin Yao, Chengjun Liu
Institution: Guangzhou Metro Design & Research Institute
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4406357813