قطعه‌بندي معنايي تصاوير معابر شهري ايران

Semantic Segmentation plays a significant role in self-driving cars, tumor detection an‎d virtual reality. However, since existing self-driving models cannot provide suitable performance for domestic roads an‎d deep-learning based models struggle with unseen images, this research seeks to introduce a neural network suited for the task of semantic segmentation of domestic images an‎d tries to minimize segmentation errors. For this purpose, a new semantic segmentation dataset called IranUrban is introduced an‎d is labeled by using CVAT, the computer vision annotation tool. Subsequently, PIDNet-L is trained on this dataset, an‎d the resulting accuracy is improved from 44.91% to 68.84% mIoU. Additionally, a novel method for improving the quality of semantic segmentation is introduced to reduce the error of unseen images. This method views the image segmentation map as a combination of polygons. Afterwards, the features of these polygons are extracted an‎d are judged by a ran‎dom forest model. Finally, the wrong polygons are removed an‎d filled with a suitable replacement. The polygons are judged based on their features to ascertain whether they should exist. It is discussed that the newly proposed method, while being simple, has a decent potential for improving the quality of semantic segmentation by utilizing decision tree models. It is seen that the implementation of the proposed method with the help of ran‎dom forest an‎d removing the wrongly detected objects improved the results of PIDNet-L, by 2 percent for motorcycle class an‎d 1 percent for person class on both CityScapes an‎d IranUrban datasets. While some classes were unaffected by the proposed method, the overall accuracy of PIDNet-L has improved by 0.27% on CityScapes an‎d 0.91% mIoU on IranUrban datasets.

5