Computational resources are limited on real-time embedded devices, so the available computing cost of deployment on the target platform must be considered. We develop a feature extraction module based on the MobileNet backbone that can be adjusted in terms of computational complexity and capacity using the depth multiplier parameter, classifier depth, and kernel depth. These three parameters allow us to control the count of channels within the network, effectively managing the model's capacity and computational requirements. To achieve semantic segmentation, we incorporate additional components, such as an extension module. This extension module typically includes 1x1 pointwise convolutional layers for pixel-level classification and a transposed convolutional layer for upsampling the output to the original input image size. By combining the feature extraction module with this extension module, we can create a complete architecture capable of performing semantic segmentation tasks. The feature extraction module provides the initial feature extraction and the extension module adds the necessary components for accurate pixel-wise classification and upsampling. Compared to Hardware-aware Neural Architecture Search (NAS), pruning, runtime pruning, and knowledge distillation methods, our model has several advantages in terms of modular design, structural controllability, ease of implementation, and cost-effectiveness. Our computational efficiency, as measured by FLOPS, is highly competitive. Our method is distinguished by solving the problem of MobileNet's inability to adjust the size and number of convolution kernels. It achieves this through adaptable parameter tuning, including MobileNet's depth multiplier, the kernel size in the FCN head's Separable Convolution layer, and the depth of the first Point-wise Convolution layer. These adjustments are customized to match hardware's max multiply-accumulates (MACs), optimizing network capacity and maximizing resource utilization.
