Research On Semantic Segmentation Method Of Complex Scene Images Based On Fully Convolutional Networ

Image semantic segmentation in complex scenes is a key issue in the direction of machine vision intelligence today.The effect of image semantic segmentation is related to the ability of visual intelligent robots or unmanned systems to understand their application scenes.The ability to accurately understand and recognize the target in the positioning scene will directly affect the judgment ability of the system and the execution efficiency of intelligent robots.With the development of intelligent vision,higher requirements are put forward for image semantic segmentation tasks in complex scenes at this stage.Although the existing semantic segmentation model based on fully convolutional neural network continuously optimizes the segmentation effect,the inherent spatial invariance of the network still leads to cause the loss of object edge details.Moreover,most models use the pixel-by-pixel loss to optimize the target,the dependencies between pixels are ignored,when facing objects with smaller spatial structures in the image,the segmentation result is not satisfactory.Therefore,this paper uses Deep Labv3+ network as the basic model to study and finally realize a complex scene image semantic segmentation method based on improved Deep Labv3+ network and superpixel edge optimization.The main research results of this paper are as follows:First of all,this paper uses group normalization instead of batch normalization and builds a Deep Labv3+ network based on group normalization to improve the performance degradation of semantic segmentation caused by setting a small batch size under the limitation of GPU memory.Replace the backbone feature extraction network with an improved Res Net50 to better adapt to semantic segmentation tasks,and adopt depthwise separable convolution in ASPP and decoder modules,reducing the amount of model calculation while ensuring accuracy.Secondly,based on the theory of relative entropy and mutual information,this paper establishes a total energy function that approximates mutual information,by modeling the dependency relationship between pixels,it can better pay attention to the structure and detailed information of small objects in space.Based on this,an overall objective loss function that integrates pixel similarity and image structure similarity is proposed and applied to the improved Deep Labv3+ network,which shows good results in the structure of small spatial targets and partial details in complex scene images.Finally,considering the special advantages of traditional image processing that superpixel segmentation retains the details of object edge information,this paper proposes a superpixel edge optimization algorithm based on simple linear iterative clustering(SLIC),which combines pixel-level semantic features and superpixel-level regional information to obtain the semantic segmentation results after edge optimization.Experiments on the PASCAL VOC 2012 dataset of multi-category complex structured objects and the real road complex scene images of the Cityscapes dataset show that the method proposed in this paper can further improve the semantic segmentation performance of complex scene images and shows better results in small target structures and object edge details.