International Journal of Circuits, Systems and Signal Processing

E-ISSN: 1998-4464
Volume 15, 2021

Notice: As of 2014 and for the forthcoming years, the publication frequency/periodicity of NAUN Journals is adapted to the 'continuously updated' model. What this means is that instead of being separated into issues, new papers will be added on a continuous basis, allowing a more regular flow and shorter publication times. The papers will appear in reverse order, therefore the most recent one will be on top.

Main Page

Submit a paper | Submission terms | Paper format


Volume 15, 2021

Title of the Paper: An Improved Object Detection Method using Feature Map Refinement and Anchor Optimization


Authors:  Yuxia Wang, Wenzhu Yang, Tongtong Yuan, Qian Li

Pages: 541-549  

DOI: 10.46300/9106.2021.15.60     XML


Abstract: Lower detection accuracy and insufficient detection ability for small objects are the main problems of the region-free object detection algorithm. Aiming at solving the abovementioned problems, an improved object detection method using feature map refinement and anchor optimization is proposed. Firstly, the reverse fusion operation is performed on each of the object detection layer, which can provide the lower layers with more semantic information by the fusion of detection features at different levels. Secondly, the self-attention module is used to refine each detection feature map, calibrates the features between channels, and enhances the expression ability of local features. In addition, the anchor optimization model is introduced on each feature layer associated with anchors, and the anchors with higher probability of containing an object and more closely match the location and size of the object are obtained. In this model, semantic features are used to confirm and remove negative anchors to reduce search space of the objects, and preliminary adjustments are made to the locations and sizes of anchors. Comprehensive experimental results on PASCAL VOC detection dataset demonstrate the effectiveness of the proposed method. In particular, with VGG-16 and lower dimension 300×300 input size, the proposed method achieves a mAP of 79.1% on VOC 2007 test set with an inference speed of 24.7 milliseconds per image.