基于深度学习的图像全景分割改进算法模型

图像语义分割和实例分割是计算机视觉领域基础挑战性工作，图像全景分割统一解决两者的任务，其核心为图像中每一个像素分配相应的类别标签以及为类别中每一个实例分配ID。经典UPSNet已经取得了较好的全景分割效果，但是使用了一种单向信息流动的特征金字塔网络，将存在实例分支的目标实例定位不够准确的问题，并且语义分支的语义分割能力还需进一步提升。本文通过考虑两个任务的差异性以及共性，重新设计特征金字塔网络结构以提取出更适合全景分割的特征图，从而提高了实例分支的AP评价指标。在语义分支中引入了克罗内克卷积，与可变形卷积进行融合使得特征图的感受野更大并且捕获了局部信息，使语义分支的mIoU评价指标得到了提高。此模型在Cityscapes数据集上进行实验，验证了所设计的每个模块及整个模型的有效性。

Improved algorithm model of image panoptic segmentation based on deep learning

Image semantic segmentation and instance segmentation are fundamental and challenging tasks in the field of computer vision. Image panoptic segmentation solves the tasks of both. Its core is to assign a corresponding category label to each pixel in the image and to assign an ID to each instance in the category. Classic UPSNet has achieved good panoramic segmentation results, but it uses a feature pyramid network with one-way information flow, which will cause the problem of insufficient positioning of the target instance of the instance branch, and the semantic segmentation ability of the semantic branch needs to be further improved . By considering the differences and commonalities between the two tasks, the paper redesigns the feature pyramid network structure to extract feature maps that are more suitable for panoramic segmentation, thereby improving the AP evaluation index of the instance branch. Kronecker convolution is introduced into the semantic branch, and the fusion with deformable convolution makes the receptive field of the feature map larger and captures local information, which improves the mIoU evaluation index of the semantic branch. The model is tested on the Cityscapes dataset to verify the effectiveness of each module designed and the entire model.