Research On Semantic Segmentation Algorithm Of Indoor Scene Based On Deep Learning

Image semantic segmentation refers to identifying the category and location of corresponding objects in an image based on semantic information,which is one of the research hotspots in the field of computer vision.Compared with the traditional image segmentation algorithm,the segmentation algorithm based on deep learning can not only automatically perform feature extraction,but also perform end-to-end training,and the segmentation accuracy and speed have been improved.However,due to the complexity of indoor scenes,semantic segmentation of indoor scenes is still extremely challenging.With the emergence and development of depth cameras,researchers have begun to use depth information to improve the accuracy of semantic segmentation.In this case,this paper studies the semantic segmentation algorithm of indoor scenes,and the main contents are as follows:1)Aiming at the problem that most of the fusion methods of semantic segmentation algorithms for indoor scenes are too single and cannot be fused according to the characteristics of RGB and depth information,a segmentation algorithm combining attention is proposed.By introducing the idea of ??attention mechanism,the feature fusion module is designed.First,the residual convolution block is used to strengthen the network's learning of the local correlation between RGB and depth information,and then the attention mechanism is used to complete the complementary fusion of RGB and depth information;at the same time,Multi-scale joint training is used to accelerate network convergence and improve segmentation accuracy.Experimental results prove that the algorithm performs well in the task of semantic segmentation of indoor scenes,and can effectively improve the accuracy of segmentation.2)In order to reduce the amount of network parameters and calculations,while further optimizing the underlying features of the image,mining depth information,In order to reduce the amount of parameters and calculations of the network,while further mining the depth information of the image,an efficient segmentation algorithm based on depth perception embedding is proposed.The algorithm first designed a feature extraction part in combination with dilated convolution,which was divided into two branches to simultaneously down-sampling and information extraction of RGB images and depth images;then,the extracted RGB feature maps were convolutional information reorganization,through cross-layer the fusion method improves the quality of low-level features and makes the features obtained by the network closer to the label.At the same time,chain residual pooling is performed on the deep feature map to enhance the network's perception of geometric information.The learned weights are used to compare the extracted geometric features.Add fusion to enhance the local relevance of geometric information;finally,an information fusion module is designed to embed depth information in stages.Experimental results show that compared with the segmentation algorithm combined with attention,this algorithm greatly reduces the amount of network parameters and calculations while ensuring the accuracy of segmentation,and effectively improves the efficiency of network segmentation.3)In order to make the algorithm in this paper meet the requirements of real-time processing,it is optimized on the basis of the segmentation algorithm based on depth-aware embedding.The number of channels of the feature map is reduced to reduce the amount of network parameters and calculations,and then the improved convolution information reorganization part is used.Make up for the detailed information and channel information lost by reducing the channel without increasing the amount of calculation too much.Experimental results prove that the optimized segmentation algorithm is superior to current mainstream real-time segmentation algorithms such as Seg Net a nd dual-stream weighted Gabor in terms of segmentation accuracy and speed.