Research On Semantic Segmentation Based On FCN With Multi-layer Information Fusion

Semantic segmentation is a key problem in image understanding.Nowadays,fully convolutional network(FCN)has become the popular model in semantic segmentation tasks.However,the repeated operations of downsampling and deconvolution in model can cause excessive loss of image information,and there are also problems such as ignoring the relationship between pixels and pixels,and lack of spatial consistency in the process of semantic segmentation.Therefore,the result of segmentation is not satisfactory,prone to blurred edges and scattered areas.Recently,many semantic segmentation methods based on FCN have been put forward,and the accuracy of semantic segmentation is also improving.But most of the semantic segmentation models based on FCN have many shortcomings,such as complex computation and too many parameters.Based on the above research status and aiming at improving the precision of image segmentation,we make a simple improvement on the basis of the FCN model,proposing two effective semantic segmentation algorithms.The first is the semantic segmentation method based-FCN with underlying contour information.The second is the semantic segmentation method based on FCN with dilated convolution and multi-layer convolution fusion.The details of the two algorithms are described as follows:(1)The semantic segmentation method based on FCN with underlying contour information overcomes the problem that the edge details of the segmentation images are too blurred caused by the multiple convolution and sampling operations in the model,which can effectively improve the semantic segmentation precision.The semantic segmentation method adds a parallel convolution branch on the basis of the FCN model,which enables the designed model to detect the underlying contour information while learning the high-level semantic features of the image.After the initial segmentation of the FCN model is completed,we use the detected underlying contour information of image to reconstruct the segmentation results.By designing multiple side-outputs,the added detection branch in the model encourages each layer to contribute to the final result rather than using the final output to modify the contour.Finally,experiments show that the contour information obtained by fusing multiple side-outputs is clearer than only using the last layer.In terms of semantic segmentation,we make a compare before and after introducing the underlying contour information,verifying the effectiveness of the design method.(2)The semantic segmentation method based on FCN with dilated convolution and multi-layer convolution fusion aims to retain more useful information by reducing the number of data scaling in the learning process,and explicitly fuses all convolution layers,thus improving the accuracy of classification prediction.The semantic segmentation model based on dilated convolution improves the traditional FCN semantic segmentation model as follows: removing the final pool of the FCN model and ensuring the invariant of the model's receptive field,the convolution operations in the fifth stage convolution layer are replaced by the dilated convolution,which reduces the multiplier of data scaling,increases the useful information of the image in the learning process,and improves the accuracy of the semantic segmentation.At the same time,this paper designs a multi-layer feature fusion structure based on the convolution layer,which explicitly fuses the features learned from all convolution layers,making up for the problem of excessive information loss in the prediction process.Finally,the feasibility and effectiveness of the algorithm is demonstrated on multiple contrast experiments.