Dual-way Network Feature Fusion Based Real-time Semantic Segmentation

Deep learning is widely used in the field of computer vision.For example,object detection,recognition,tracking,scene understanding,etc.Among them,image semantic segmentation has always occupied an important position.The full convolutional neural network exhibits an unprecedented advantage in dealing with image semantic segmentation.The most commonly used full convolutional network is the Encoder-Decoder network.During the encoding process,the original image is downsampled,so a large amount of spatial information is lost,resulting in reduced segmentation accuracy.In addition,the usual algorithm model will adopt a relatively large neural network structure,resulting in a decline in the speed of reasoning.However,if the lightweight network is simply used as the encoder,for efficient consideration,the network's receptive field will drop significantly,resulting in a decrease in the segmentation accuracy.To solve the above problems,this thesis proposes a real-time semantic segmentation algorithm model based on dual way network and different hierarchical feature fusion.The main innovations are as follows:(1)We improve the existing lightweight network by combining the advantages of both deep separable convolution and hole convolution which effectively increase the receptive field of lightweight network and the ability to extract multi-scale features,and improve the lightweight The quality of the features extracted by the network encoder.(2)We propose the attention mechanism module to filter the features extracted by the encoder to improve the quality of the features maps.We also extract the upper and lower information of the image at the same time,which further improves the segmentation accuracy.(3)An efficient feature fusion method is proposed,which efficiently fuses the spatial information and the semantic information of the image.In this way,we solve the problem of space information lacking problem in the high-level features.The two features complement each other and improve the performance of the algorithm.The proposed algorithm is validated on two semantic segmentation datasets.Compared with other real-time semantic segmentation methods,our method achieves the best segmentation performance,realizing high accuracy and real-time semantic segmentation simultaneously,and successfully migrates to Android mobile device.