Research On Semi-supervised Semantic Segmentation Algorithms Based On Deep Learning

In recent years,image semantic segmentation has emerged in many fields with the widespread use of deep learning in the field of image semantic segmentation.It is often applied in practical scenarios such as autonomous driving,geological exploration and urban planning,helping people to improve productivity and promote scientific research.Semantic segmentation algorithms based on deep learning solve the problem of mis-segmentation in complex backgrounds and the difficulty of handling complex object shapes in traditional image segmentation.Compared with traditional image segmentation methods,existing semantic segmentation methods can show higher segmentation accuracy by learning features after training a large number of images,and do not require manual design of feature rules.Therefore,deep learning-based semantic segmentation methods are of great research importance.At the same time,the labeling of training images requires manual annotation,which wastes a lot of human and material resources.With the development of semi-supervised semantic segmentation,this problem is gradually solved,i.e.the semantic segmentation task can be completed by using only part of the training with labeled images.Currently,semi-supervised semantic segmentation models or algorithms based on deep learning have problems such as high model complexity and demanding hardware equipment requirements.To solve these problems,this thesis proposes a semi-supervised semantic segmentation model based on adversarial learning using an adversarial learning training strategy,which effectively improves the efficiency and model accuracy,respectively,and has obvious effects.The innovation points are as follows:(1)Using the adversarial training scheme,the generative network in the generative adversarial network is replaced by a segmentation network with weight normalised convolution as the core,using a control smoothing effect to transfer from activation to weights,making the gradient of the model stable and more efficient to train,improving the training efficiency by 47.3% compared to existing methods.(2)As the convolution and self-attentiveness mechanisms have a lot of repeated calculations for information in the same region during feature aggregation,this thesis invokes a 1×1 convolution,using intermediate features for different aggregation operations to reduce the repeated calculations of convolution and self-attentiveness and to reduce the model complexity and computational redundancy.(3)In this thesis,a lightweight location-aware circular convolutional architecture is used as the backbone network of the model for obtaining deep feature information.This method can obtain structural information over a larger area,and the computational process is simple and less complex,which can further reduce the number of parameters in the model to 38.6M.(4)In order to enhance the feature representation capability of the model and reduce the phenomenon of mis-segmentation,the model incorporates a dual spatial and channel polarization attention mechanism to achieve complete collapse of the features in the other direction while keeping the loss of compressed information in the orthogonal direction small,for the purpose of improving segmentation performance.The model designed in this thesis is compared with existing state-of-the-art semi-supervised semantic segmentation models,and the experimental results on publicly available datasets show that the segmentation accuracy is improved to 76.32%,and the training time is only 1/10 of that of the compared state-of-the-art models.the number of model parameters is lower,and a good balance between training efficiency and segmentation accuracy is achieved.