Fuzzy-Logic based Effective Contour Representation of Occluded Objects

We present a fuzzy-based network for the sharpening of object contour even in the presence of occlusion. The contour representation of objects can be effectively handled by the structure tensor method. This work proposes an occlusion detection and filling strategy using the square patch selection method. Based on the interpolation method, the fuzzy-assisted square patch selection can be used to fill the occluded pixels. Due to the occluded pixels, the depth map may have anomalies in the low-texture and high-exposure areas. Before converting a depth map to a point cloud, it is essential to filter out the outliers in the depth map to obtain a more accurate point cloud. To improve the precision of the depth map, improved occlusion detection and management procedures is required.The occlusion regions may be confirmed through belief propagation, which may produce noisy results in occluded regions, sharp objects, and object boundaries. We strived to build a model that differentiates the occluded pixels from others by exploiting sharp boundary transitions. We have used a stereo geometry structure to develop the required deep neural models to handle occlusion. We built the model by creating layers for every pipeline component and made it to learn the contour representation model using an adaptive fuzzy-based approach. In existing approaches, the bias must be properly predicted with the Gaussian distribution. The proposed model eradicated the pixel bleeding effect by exploiting the bimodal distribution with Gaussian and SMD (Stereo Mixture Density) functions and by finding smoothening bias.The suitable depth values were assigned to the occluded regions obtained. The experimental results demonstrated that the proposed approach generates more stable depth maps with fewer constraints than the existing methods. The experimental results were compared with the standard SMD-Net and other state-of-the-art models.