DWT-Based ResNet34 Framework for Road Segmentation

Abstract

Road segmentation is a key task in intelligent vehicle systems, autonomous driving, and urban scene understanding. In this work, we propose a DWT-based ResNet-34 framework for road segmentation that uses the Discrete Wavelet Transform (DWT) and ResNet-34 to extract road regions from the Cityscapes database.  First,  resize and normalize the original Cityscapes images,  then apply a one-level 2D-DWT using the Haar wavelet to decompose the image into four sub-bands: LL, LH, HL, and HH. The LL-band is selected as input to the ResNet-34-based segmentation model because it retains important low-frequency structural information while reducing unnecessary high-frequency noise. The ResNet-34 encoder extracts hierarchical spatial features from the LL-band image through convolution, max pooling, and residual blocks. The final encoder feature map is passed to a decoder/upsampling block to generate the final road segmentation mask. The proposed DWT with the ResNet-34 method is evaluated using typical performance metrics, including accuracy, precision, recall, F1 Score, and Intersection over Union (IoU). The model achieved 93.03% accuracy, 79.57% precision, 87.66% recall, an F1-score of 83.42%, and an IoU of 71.55%. The outcomes show that using the LL-band from DWT improves road segmentation performance by enhancing salient structural information and suppressing irrelevant image details. Therefore, the proposed DWT-based ResNet-34 model provides an effective approach for road segmentation in urban scene images.