| 摘 要: 为应对DeepLabV3+在图像语义分割中存在的资源消耗较大与边缘特征信息损失问题,设计了一种兼顾计算效率与分割精度的小目标场景下多尺度轻量级语义分割模型。引入MobileNetV2作为骨干网络以压缩参数规模,降低整体模型的复杂度;在浅层特征路径中嵌入混合多尺度注意力(HMSA)机制,提升模型对边缘与纹理特征的敏感性;同时在深层特征中融合坐标注意力(CA)机制,以强化语义表达与位置信息的协同建模;此外,优化损失函数结构,结合CE与Dice损失函数,提升模型在样本不均衡场景下对小目标的关注度与分割表现。改进模型在PASCALVOC2012数据集和自建山桃花语义分割数据集(GHFSD)上进行实验,较原模型平均交并比(mIoU)分别
提升3.14和4.19个百分点,且模型参数量仅增加0.025×106。综合实验结果表明,所提出的集成注意力机制与多尺度特征增强策略的轻量级语义分割模型,在保持分割精度的同时,显著降低了计算资源需求,具备在边缘设备及高精度分割任务中的实际应用潜力。 |
| 关键词: 语义分割 轻量化网络 注意力机制 多尺度融合 目标提取 |
|
中图分类号:
文献标识码: A
|
|
| Multi-scale Lightweight Semantic Segmentation Model Optimized for Small-target Scenarios |
|
YANG Qihang, LIU Liqun
|
(College of Information Science and Technology, Gansu Agricultural University, Lanzhou 730070, China)
yangqh@st.gsau.edu.cn; llqhjy@126.com
|
| Abstract: To address the challenges of high computational resource consumption and edge feature information
loss in DeepLabV3 + for image semantic segmentation, this paper proposes a mult-i scale lightweight semantic segmentation model optimized for smal-l target scenarios that balances computational efficiency and segmentation accuracy. The proposed solution first employs MobileNetV2 as the backbone network to reduce parameter scale and overall model complexity. A Hybrid Mult-i Scale Attention (HMSA) mechanism is embedded in the shallow feature pathway to enhance the model’s sensitivity to edge and texture features. Simultaneously, a Coordinate Attention (CA) mechanism is integrated into deep features to strengthen the collaborative modeling of semantic representation and positional information. Additionally, the loss function structure is optimized by combining CE and Dice loss, im_x005f_x0002_proving the model’s attention toward small targets and segmentation performance under class imbalance scenarios.Experimental results on the PASCAL VOC 2012 dataset and a sel-f built mountain peach blossom dataset (GHFSD) demonstrate that the improved model achieves mean Intersection over Union (mIoU) improvements of 3.14 and 4.19 percentage points respectively compared to the original model, with only a 0. 025-million increase in parameters.Comprehensive experiments indicate that the proposed lightweight segmentation model, incorporating integrated attention mechanisms and mult-i scale feature enhancement strategies, significantly reduces computational resource demands while maintaining segmentation accuracy, showing practical application potential for edge devices and high-precision segmentation tasks. |
| Keywords: semantic segmentation lightweight network attention mechanism multi scale fusion object extraction |
