| 引用本文: | 杨海燕1,2 ,王凤随1,2 ,张兴旺1,2.基于上下文增强和特征融合的目标检测算法(J/M/D/N,J:杂志，M：书，D：论文，N：报纸).期刊名称,2025，42（3）：102-109 |
| CHEN X. Adap tive slidingmode contr ol for discrete2ti me multi2inputmulti2 out put systems[ J ]. Aut omatica, 2006, 42(6): 4272-435 |
|
| 摘要: |
| 目的 针对 CenterNet 无锚框目标检测算法表征能力不足的问题,提出一种基于上下文增强和特征融合的学
习方法。 方法 该方法采用多感受野和信息融合的思想,构建自适应上下文提取模块和特征融合策略。 首先网络通
过自适应上下文提取模块的多路径空洞卷积层获取目标的上下文特征,督促深层网络学习多尺度信息;然后,通过
ACON-C 激活函数在网络中加入非线性因素,对网络神经元自适应地激活,增强网络的数据拟合能力;最后,联合
注意力特征融合策略对不同层次的特征信息进行合并,通过整合深层网络的语义信息和浅层网络的位置信息,来
捕获对识别任务有用的特征信息,同时学习特征图在多个层次通道间的相关性,以加强网络对关键目标特征的专
注度。 结果 所提方法在 PASCAL VOC 公开数据集上 mAP 达到 83. 82%,约比基线算法 CenterNet 增加了 3. 72%。
相较于经典算法 Faster R-CNN、SSD、YOLOv3 分别增加了 7. 4%、9. 5%、3. 5%。 结论 有效地提升了 CenterNet 算法
的检测性能,并且改进的 CenterNet 相较于其他目标识别算法具有更高的识别准确度,在目标检测应用中具有良好
的实用性,充分验证了所提方法的有效性。 |
| 关键词: 目标检测 空洞卷积 上下文特征 特征融合 注意力机制 ACON 激活函数 |
| DOI： |
| 分类号: |
| 基金项目: |
|
| Target Detection Algorithm Based on Context Enhancement and Feature Fusion |
|
YANG Haiyan1 2, WANG Fengsui1, 2 ZHANG Xingwang1 2
|
|
1. School of Electrical Engineering Anhui Polytechnic University Anhui Wuhu 241000 China
2. Key Laboratory of Advanced Perception and Intelligent Control of High-end Equipment Ministry of Education Anhui
Wuhu 241000 China
|
| Abstract: |
| Objective Aiming at the issue of inadequate characterization capacity of CenterNet anchor-free object
recognition algorithm a method based on context enhancement and feature fusion was proposed. Methods This method
adopted the concepts of multi-receptive field and information fusion to construct an adaptive context extraction module and
feature fusion strategy. Firstly the network obtained the contextual features of the target through the multipath dilated
convolution of the adaptive context extraction module prompting deep networks to learn multi-scale information. Then
nonlinear factors were added to the network through the ACON-C activation function adaptively activating the neurons of the network and enhancing the data-fitting ability of the network. Finally a joint attention feature fusion strategy was used
to merge feature information at different levels. By integrating semantic features from the high-level network and positional
features from the low-level network the feature information that is useful for the recognition task was captured. At the
same time the correlation between the feature maps in multiple levels of channels was learned to enhance the network?? s
focus on key target features. Results The proposed method achieved an mAP of 83. 82% on the PASCAL VOC public
dataset an improvement of 3. 72% compared with the CenterNet baseline algorithm. It also outperformed classic
algorithms such as Faster R-CNN SSD and YOLOv3 by 7. 4% 9. 5% and 3. 5% respectively. Conclusion The
proposed method effectively enhances the detection performance of the CenterNet algorithm and the improved CenterNet
has higher recognition accuracy compared with other target recognition algorithms. The proposed method proves to be
practical in target detection applications validating the effectiveness of the proposed approach. |
| Key words: target detection dilated convolution context feature feature fusion attention mechanism ACON activation
function |
