Fig 1.
The structure diagram of residual.
Residual network uses jump connection to construct residual structure, which solves the degradation problem to some extent.
Table 1.
ResNet-101 network structure table.
Fig 2.
The structure diagram of DeppLabv3+.
The Deeplabv3 + adopts residual network as the backbone network, and uses encoding and decoding structure to improve semantic segmentation effect [18].
Fig 3.
The structure diagram of Res2Net.
Fig 4.
Extended visual field of receptive field.
Fig 5.
Optimization structure diagram of multi-loss constraint.
Fig 6.
Flow chart of semantic segmentation of urban road images.
Table 2.
Overall segmentation results of Cityscapes (unit: %).
Table 3.
Segmentation results of Cityscapes dataset experiment (unit: %).
Table 4.
Overall segmentation results of CamVid dataset (unit: %).
Table 5.
Segmentation results of CamVid dataset experiment (unit: %).
Fig 7.
Partial segmentation results of Cityscape dataset.
Comparing the first column of images, it can be seen from the labeled area of the yellow box in the figure that the truck head in the original algorithm segmentation result is mis-segmented, and the segmentation results of the Improved Ⅰ and Improved Ⅱ algorithms reduce the mis-segmented area. After applying two innovative methods at the same time, the segmentation area is minimized. In the second column of image labeling area, the original, Improved Ⅰ, and Improved Ⅱ algorithm segmentation results have the phenomenon of mis-segmentation of cars into trucks, and the proposed algorithm is more correct to segment the contours and categories of cars. In the third column of segmentation results, the bicycle contours in the original algorithm and the Improved Ⅰ are relatively fuzzy. In addition, there is a problem of discontinuous segmentation of street lights, and the proposed algorithm can get a clear bicycle contour shape. In the fourth column of images, it can be seen that part of the truck area in the original algorithm was mistakenly divided into cars and sky categories. The Improved Ⅰ and the Improved Ⅱ algorithm reduced the mis- segmented area, and the proposed algorithm segmented the truck area completely and correctly.
Fig 8.
Results of partial segmentation of CamVid dataset.
From the image segmentation results in column (a), it can be seen that the Deeplabv3plus algorithm, the Improved Ⅰ and the Improved Ⅱ have different degrees of under-segmentation on the street light poles, and the complete street light pole area cannot be segmented. Compared with the above three methods, the proposed algorithm can segment a more complete street light pole area and get more precise details. In view of the images in column (b) and their segmentation results, Deeplabv3plus, Improved Ⅰ, and Improved Ⅱ algorithms have different degrees of mis-segmentation in the segmentation of the motor vehicle lane area, and part of the motor vehicle lane area is mistakenly divided into sidewalks. In practical applications, the misclassification of these two semantic categories is a fatal error. The algorithm of this paper, which combines two improved strategies, correctly divides the motor vehicle lanes and sidewalks, and solves the mis-segmentation phenomenon. For the images in column (c) and their segmentation results, the sidewalk on the left and the street lights on the right have different degrees of mis-segmentation, especially in the segmentation results of Deeplabv3plus and the improved algorithm. Although the Improved Ⅱ segmented the sidewalk area relatively correctly, the streetlight area marked on the right side was seriously under-segmented. The segmentation result of the algorithm in this paper is still relatively rough compared to the label, but compared with the above segmentation result, the error segmentation phenomenon is reduced. In the image (d) and a series of segmentation results, Deeplabv3plus and the Improved Ⅰ algorithm are not accurate in the segmentation of the marked sidewalk area, while the Improved Ⅱ algorithm and our algorithm obtain more accurate segmentation results.