Abstract: Superpixels exhibit superior image representation capabilities and exceptional computational efficiency, making them widely adopted in downstream computer vision tasks. However, existing superpixel segmentation algorithms insufficiently account for the coupling interplay between semantic and spatial cues, resulting in suboptimal spatial coherence and semantic consistency of the generated superpixels and leading to erroneous delineation of fine details in complex scenes. To address these shortcomings, this paper proposes superpixel segmentation algorithm based on Inverted Feature Pyramid Network (IFPNet), an end-to-end architecture that fuses multi-scale semantic context with spatial detail through a cascaded coupling of an encoder and an inverted feature pyramid decoder. In addition, we introduce a Perceptual Boundary-Driven Composite Loss function to further enforce semantic consistency. Comprehensive experiments on the BSDS500, NYUv2, and KITTI datasets and compared with state-of-the-art (SOTA) algorithms (i.e., SLIC Superpixels Compared to State-of-the-Art Superpixel Methods (SLIC), Superpixel Sampling Networks (SSN), and Superpixel Segmentation with Fully Convolutional Networks (SCN)), demonstrate that IFPNet achieves state-of-the-art performance.: on BSDS500, it attains an Achievable Segmentation Accuracy (ASA) of 0.9720 and a Boundary Precision (BP) of 0.1306; on NYUv2, Achievable Segmentation Accuracy (ASA) and Boundary Precision (BP) reach 0.9482 and 0.1999, respectively; and on KITTI, Achievable Segmentation Accuracy (ASA), Boundary Precision (BP) and Compactness (CO) improve to 0.9655, 0.1470, and 0.3720. The experimental results confirm that the proposed IFPNet generates superpixels with enhanced spatial coherence, semantic fidelity, and superior preservation of detailed texture, underscoring the robustness and generalizability of IFPNet.