Page 21 - JCAU-6-1

Journal of Chinese
Architecture and Urbanism Spatial exploration through image semantic segmentation

Figure 3. Partial image stitching of Street 1. Source: Images by the author
was captured and saved separately every 10 s. Subsequently, 2.4.3. Semantic segmentation and subsequent
the captured frames underwent brightness adjustment and processing of street images of Lu Xun’s hometown
noise reduction to address instances of excessive darkness The pre-processed street images from Lu Xun’s hometown
and noise, thereby facilitating the subsequent processing. were incorporated into the abovementioned training
(b) Normalization dataset for semantic segmentation. This process resulted
in the generation of segmented color block images. The
The resolution dimensions of the 279 images were obtained results of semantic segmentation are exclusively affected by
through the abovementioned process and normalized by Python. the proportion of different color block regions. To reduce
2.4. Street scene analysis based on semantic computational workload, further normalization processing
segmentation was applied to the color block images. During this process, the
image resolution was compressed to facilitate pixel statistics.
2.4.1. Selection of semantic segmentation model
2.4.4. Pixel-wise image analysis using Python
The semantic segmentation process utilized a classic
CNN model, ResNet-22, implemented through CUDA- In this step, Python was utilized to scan images pixel by
accelerated computation, with the training framework pixel, tallying and merging pixels belonging to various
facilitated by Baidu’s PaddlePaddle. Cityscape trained color groups. Statistical charts were subsequently generated
the datasets using a series of high-resolution images based on the counting results. The results obtained from
encompassing various urban scenes, including roads, various dimensions were aggregated to calculate quantities,
buildings, pedestrians, and others. The extensive dataset, averages, variances, and other parameters, facilitating
consisting of thousands of high-precision annotated subsequent analysis (Figures 5 and 6).
training images, was methodically divided into training, 2.5. Comparison and analysis of different street
validation, and test sets. The accuracy of well-trained spatial features in Lu Xun’s hometown
Cityscape images could reach up to 70% or even 80%.
2.5.1. Spatial features of the historic street (Street 1)
2.4.2. Model network architecture and training process in Lu Xun’s hometown

Eight pre-defined labels were adopted in Cityscape, namely The distribution of various elements in historic streets
road, vegetation, building, sky, person, vehicle, wall, and sign. is worthy of examination. As illustrated in Table 1, road

Volume 6 Issue 1 (2024) 4 https://doi.org/10.36922/jcau.1736