Research On Image Semantic Segmentation Method Based On Context Feature Fusion

Image semantic segmentation is one of the most important research directions in the field of computer vision.Semantic segmentation is to assign pixel-level labels to objects in an image,and accurate identification and accurate segmentation of objects is crucial when facing highprecision,high-security tasks.The target space faced by semantic segmentation is huge and complex.For different images,the perspective relationship,illumination intensity,scale,and color complexity of various objects are different.The complexity of the target space generally causes boundary distortion in the segmentation of small-scale targets and the local boundary is discontinuous.In view of this,this paper proposes a new image semantic segmentation method based on context feature fusion to effectively solve this problem.question.The method is trained,validated and tested on the public dataset VOC2012.The proposed method improves the acquisition ability of local features,enables to obtain finer local feature information while segmenting large objects well,and improves the segmentation accuracy of local object boundaries.The main research results are as follows:(1)An encoder architecture utilizing multi-scale contextual feature fusion is designed.The overall semantic segmentation network architecture draws on the encoder-decoder structural characteristics of DeepLab V3+,combined with the idea of positive correlation between the receptive field scale and eccentricity in RFBNet,and adopts the atrous convolution stacking processing of different hole rates to form a multiscale stacking context feature fusion module,using this module to obtain receptive fields of different sizes is conducive to extracting detailed information distributed in different sizes,and improves the segmentation accuracy of small-scale targets.(2)Design a decoder architecture that utilizes local contextual feature fusion.Based on the multi-scale context feature fusion encoder architecture,an improved subpixel convolution is used in the decoder structure to replace the original bilinear upsampling operation,and the resolution of the feature map is restored.The idea of reordering channel information further improves the overall segmentation accuracy of the network and trains the final model.Through a large number of experimental comparative analysis,the accuracy of the method in this paper on the VOC2012 dataset is 86.44%mIoU,which is 3.99%higher than when only the multi-scale context feature fusion encoder architecture is used.Compared with CaC-Net,DMNet,PSPNet and Semantic segmentation models such as DeepLab v3+are 1.34%,2.04%,3.84%,and 2.86%higher,respectively.It is 0.84 percent higher than the state-of-the-art algorithm,and the experimental results demonstrate the effectiveness of our method.Finally,the training test is performed on the Plant Semantic Segmentation dataset,and the segmentation accuracy reaches 95.7%mIoU.The results show that the method in this paper has good generalization performance.