SBSN: Weakly Supervised Image Segmentation Based on Shape Prior

In the domain of multimedia signal processing, shape priors serve as crucial cues for guiding model training. However, many existing techniques struggle to effectively represent complex shape priors in multimedia content. To address this challenge, we introduce a Shape-Based Image Segmentation Network (SBSN) that leverages a Convolutionally Constrained Boltzmann Machine (CRBM) for shape prior extraction from multimedia signals. The SBSN employs CRBM to explicitly model the shape probability distribution through rigorous mathematical formulations and utilizes Wasserstein distance as a metric to quantify the disparity between the segmentation output and the extracted shape prior. For multimedia signal segmentation, SBSN incorporates an advanced feature fusion architecture, utilizing ResNet for hierarchical feature extraction and transposed convolution for signal upsampling and reconstruction. This process culminates in high-resolution, pixel-level segmentation of multimedia content. Notably, SBSN is designed to operate effectively with a limited set of labeled multimedia samples during the training phase. Extensive evaluations were conducted against state-of-the-art weakly-supervised methods across diverse multimedia objects: human figures, horses, airplanes, and surface cracks. The results demonstrate that SBSN achieves a significant 0.8% improvement in average AUROC and secures the second rank in average ABO performance, underscoring its efficacy in processing and analyzing complex multimedia signals.