Research On Image Semantic Segmentation Network Based On Multi-scale Perception And Attention Mechanism

Semantic segmentation is a technique which classifies and labels the objects of a picture pixel by pixel according to the definitions of language.For example: the sky,grass,and people in the image are marked with blue,green,and red respectively.As the upstream task of image analysis,semantic segmentation is related to human life closely and it is widely used in automatic driving,geographic information integration,medical image diagnosis and so on.Most of the existing algorithms of semantic segmentation are implemented by neural networks.the task of image segmentation in complex scenes has always been a difficult task in this field.semantic segmentation in natural scenes is affected by the diversity of object's shape,distance,illumination and other factors easily.We proposes a new dualbranch semantic segmentation network based on multiscale strip pooling and attention mechanism.The main contributions are summarized as follows:1.By introducing one-dimensional atrous convolution and multi-scale in the subnet of spatial-aware to optimize the strip pooling model.this operation can further enlarge the receptive field of the model in the horizontal and vertical directions,the strip pooling technology is based on the strip pooling operators(the kernel size is n×1).compared with the commonly used square(n×n)size of kernel,the strip pooling can enlarge the perception of large-scale objects in the image in the horizontal and vertical directions without adding the extra parameters.2.Using the pre-trained VGG16 as the content-aware subnet to assist the subnet of spatial-aware to optimize the embedded features of image semantic segmentation.And complete the feature fusion based on the subnet of content-aware and spatial-aware to enhance the characterization ability of feature maps.3.Use the SOCA(second-order channel attention)to optimize the feature selection of the middle-level and high-level.By using the strategy promote the important channel information and suppress irrelevant channel information in the process of training to improve the quality of segmentation;We use covariance and gating mechanisms to achieve the model of SOCA to enhance the expression of the feature.Extensive experiments on popular benchmarks(Cityscapes)demonstrate that our approach get a great result and the metrics of segmentation m Io U is improved by 1.2%;the Ablation Studies show that the improved strip pooling architecture,the second-order channel attention,and the content-aware auxiliary network played a positive role on the improvement of the experimental results.