Research On Semantic Segmentation Algorithm Based On Deep Learning

Image semantic segmentation is one of the core tasks of computer vision,and its purpose is to effectively classify each pixel of the input image.In recent years,deep learning is the most far-reaching technology in the computer field.With the help of deep learning,image semantic segmentation tasks have achieved many results in the areas of autonomous driving,biomedicine,and augmented reality.Compared with image classi-fication and object detection,semantic segmentation can provide richer image semantic information.However,there are many problems in current deep learning-based semantic segmen-tation.Firstly,the semantic segmentation data set is difficult to produce,and it has the problems of training difficulty and high production cost.Secondly,most of the algorithm computation and parameters are huge,which makes it impossible to apply to mobile de-vices with limited computing resources,limiting the development of semantic segmen-tation.In addition,many algorithm applications do not make full use of the hardware resources of the computing platform to accelerate the program running speed.Therefore,this paper mainly researches and optimizes these three aspects.The main contents and innovations are as follows:1.Efficient Train.Through a detailed analysis of the existing weak supervision algo-rithms,this paper proposes the RGrad-CAM algorithm,which uses image classification-level labels to train the network and outputs a heat map to achieve efficient network train-ing and greatly reduces training costs.A detailed visual analysis of the algorithm is carried out,and the reasons of the improvement is thoroughly explored from three aspects:heat map,feature map and gradient map.RGrad-CAM improves the accuracy of the heat map by promoting the weight of some feature maps.The segmentation test results on the PAS-CAL VOC dataset show that the mIoU index of RGrad-CAM is 3%higher than the CAM,and other indicators are also better than it.2.Efficient Model.This paper proposes an efficient semantic segmentation network,called EEDNet,based on the encoder-decoder structure.The EEDNet uses MobileNet as the encoder;it uses the attention mechanism to achieve efficient feature extraction and dimensionality reduction,reducing the overall computation;making full use of the de-coder's classification output results,and using its rich context information to assist in segmentation,thereby Improve segmentation accuracy;high-level feature maps help low-level feature maps to recover their spatial semantic information,and multi-layer feature maps are effectively fused.Experimental results on multiple standard data sets show that EEDNet achieves a very good balance between segmentation accuracy and efficiency.3.Efficient Inference.With the same time and space complexity of the algorithm,this paper makes full use of the computer architecture to optimize the program's operating efficiency.Adopt the advantages of low precision in memory layout and computation efficiency to speed up program running speed.Use existing GPU engines to accelerate the inference.Experimental results show that low accuracy can significantly speed up the inference without affecting the accuracy,and is a very valuable engineering optimization solution.