基于多尺度跨层特征融合的轻量化小目标检测算法

Lightweight Small Target Detection Algorithm Based on Multi-scaleCross-layer Feature Fusion

针对现有算法缺乏对小目标检测算法的优化且存在计算资源浪费的问题,提出基于MobileNetV2-SSDLite的轻量化小目标检测算法。首先在骨干网络MobileNetV2中嵌入坐标注意力机制,聚焦图像中感兴趣的区域,提出一种基于PANet的多路径、跨层级的特征交叉融合网络,以改善低级特征图的空间语义信息,增加全局信息的捕获。在路径中部署扩张感受野模块获取更多浅层特征的上下文信息,增强小目标检测能力。在PASCAL VOC2007上对该方法进行定量和定性分析,实验结果表明,相对于基线算法准确度提升了4.5%,且在小目标类别上的检测效果也显著提升,能够适应室内外场景中密集小目标的检测任务,满足轻量化模型对小目标检测的需求。

Aiming at the problem that existing algorithms lack the optimization of small target detection and there is a waste of computing resources,a lightweight small target detection algorithm based on MobileNetV2-SSDLite was proposed.Firstly,coordinate attention mechanism was embedded in the backbone network MobileNetV2 to focus on interesting region in the image.A multi-path and cross-level feature cross-fusion network based on PANet was proposed to improve spatial semantic information of low-level feature map and increase the capture of global information,deploy the expanded receptive field module in the path to obtain more shallow feature context information and enhance the small target detection ability.The quantitative and qualitative analysis of the method was carried out on PASCAL VOC2007,and experimental results showed that method accuracy was improved by 4.5 percentage points in comparison with the baseline algorithm.In addition,the detection effect on small target categories was also significantly improved,which could adapt to detection task of dense small targets in indoor and outdoor scenes,and met the requirements of lightweight model for small target detection.