基于混合注意力及自适应多尺度的语义分割算法研究

Research on Semantic Segmentation Algorithm Based on Hybrid Attention and Adaptive Multiple Scales

为提升语义分割网络特征提取的有效性以及尺度不变性,提出了一种基于混合注意力机制和多尺度特征自适应融合的轻量级语义分割算法。算法采用颜色空间转化、边缘提取以及灰度化等图像预处理方法增强输入图像信息;利用深度可分离卷积、池化和H-Swish激活结合残差结构逐步提取目标局部和全局特征,并设计混合注意力机制从最大、均值和标准差等角度分别捕获特征通道及空间位置的全局上下文信息,使网络聚焦目标关联特征,降低背景信息干扰;针对不同大小目标,引入了多尺度特征自适应加权融合结构,以自主选择的方式来避免各尺度目标特征相互影响。通过在标准、仿真以及实际场景数据集上的实验结果表明,所提方法有效提升了特征多样性以及关键特征的贡献,保障了多尺度目标准确识别,并能较好地应用于实际场景中,高效实现语义分割任务。

In order to improve the efficiency of feature extraction and scale invariance of semantic segmentation network,a lightweight semantic segmentation algorithm based on hybrid attention mechanism and multi-scale adaptive feature fusion is proposed.Firstly,image preprocessing methods such as color space transformation,edge extraction and graying are used to enhance the input image information;Secondly,depth-wise separable convolution,pooling,H-Swish activation combined with residual structure are applied to gradually extract local and global features of the object,and a hybrid attention mechanism is designed to capture the global context information of feature channel and spatial position from the perspectives of maximum,mean and standard deviation,so that the network can focus on the object related features and reduce the interference of background information.Finally,for objects of different sizes,a multi-scale feature adaptive weighted fusion structure is introduced to avoid the mutual influences of object features of each scale by means of autonomous selection.The experimental results on standard,simulation and actual scene datasets show that the proposed method effectively improves feature diversity and the contribution of key features,ensures accurate multi-scale object recognition,and can be better applied to actual scenes to implement semantic segmentation tasks efficiently.