Research On Semantic Segmentation Of Cityscape Based On Deep Neural Network

Artificial intelligence technology has been rapidly developed,the term “perceive city” has gradually attracted attention and familiarity.Using image semantic segmentation technology to “perceive city” is an efficient and low-cost way.For example,it helps the pilot to determine the safe landing position by segmenting aerial image of the aircraft.It helps the driver to plan safe and reliable driving routes by segmenting the images transmitted by the driving recorder.In recent years,deep learning method has made a series of breakthroughs in semantic segmentation.However,in practice,segmentation models tend to be too large in parameter size and poorly portable,which is not conducive to practical engineering applications.Meanwhile,the segmentation results are usually not ideal due to complex cityscape scenes.In view of the above problems,this paper deeply studies the image semantic segmentation method under cityscape scenes.The main works are as follows:From the perspective of model compression,we propose a lightweight multi-channel feature fusion semantic segmentation method aming at complex cityscapes scenes segmentation.Firstly,we improves on Deeplab V2 segmentation model,and combines the Deeplab V2 with the multi-channel feature fusion structure Refine Net to effectively improve the segmentation details.Then,based on it,we made some adjustments to the model.We introduce depthwise separable convolution into ASPP and Refine Net and propose LightASPP and Light-Refine Net V1 and V2,which effectively reduces the large number of parameters.The experimental results show that the lightweight multi-channel feature fusion semantic segmentation model we proposed can obtain better segmentation results with less parameter,which can meet the general practical application.In order to extract the comprehensive and effective feature in the image and further improve the segmentation results of cityscape,we propose a feature fusion model based on deformable convolution.On the encoder,we use the lighter CNN model Xception as the base model,and use the ASPP to extract multi-scale features to have a better segment performance of different size targets.On the decoder,we propose a feature fusion decoder based on deformable convolution.We introduces the deformable convolution into the decoder structure for more accurate extraction of deep and low level CNN features,and further fine-tunes the merged features with the chained residual pooling module to obtain more accurate segment the result.The experimental results show that the feature fusion segmentation model based on deformable convolution we proposed can accurately segment cityscape images and improve segmentation details.In summary,the semantic segmentation method in cityscape proposed in this paper can achieve good results in both practical applications and high-precision image segmentation,and has strong scalability.