Deep Learning-Driven Approaches for Optimizing Accuracy and Inference Speed in Compact Segmentation Models on Edge Devices