基于DeepLabV3+模型改进的图像分割方法研究

Research on Improved Image Segmentation Method Based on DeepLabV3+Model

近年来,计算机视觉快速发展,其中图像分割在计算机视觉中也举足轻重,在城市现代化建设,智能驾驶,地理勘测等方面都得到了充分应用。但是,大多数分割方法只关注图像特征的纵向深层特征与浅层特征的简单融合,而忽略了同一层图像特征的横向远程关系。针对此问题,基于DeepLabV3+框架,加入Swin-Transformer block,利用其自注意力机制特点,进行网络特征提取,以提高图像分割的全局和细节优化。其次,改进DeepLabV3+中上采样方法,利用CARAFE上采样模块取代简单的双线性插值法。实验表明,改进后的模型相较于基线模型MIoU提升2%,ACC提升1%。

In recent years,Computer Vision has developed rapidly,in which image segmentation also plays a decisive role in Computer Vision,and has been fully applied in urban modernization construction,intelligent driving,geographic survey,and so on.However,most segmentation methods only focus on the simple fusion of vertical deep features and shallow features of image features,while ignoring the horizontal remote relationship of image features in the same layer.To address this problem,based on the DeepLabV3+framework,the Swin-Transformer block is added,and its Self-Attention Mechanism feature is utilized for network feature extraction in order to improve the global and detailed optimization of image segmentation.Secondly,the up-sampling method in DeepLabV3+is improved,and CARAFE up-sampling module is utilized to replace the simple Bilinear Interpolation method.Experiments show that the improved model increases MIoU by 2%and ACC by 1%compared with the baseline model.