Search

Semantic segmentation has been a major topic in research and industry in recent years. However, due to the computation complexity of pixel-wise prediction and backpropagation algorithm, semantic segmentation has been demanding in computation resources, resulting in slow training and inference speed and large storage space to store models. Existing schemes that speed up segmentation network change the network structure and come with noticeable accuracy degradation. However, neural network quantization can be used to reduce computation load while maintaining comparable accuracy and original network structure. Semantic segmentation networks are different from traditional deep convolutional neural networks (DCNNs) in many ways, and this topic has not been thoroughly explored in existing works. In this paper, we propose a new quantization framework for training and inference of segmentation networks, where parameters and operations are constrained to 8-bit integer-based values for the first time. Full quantization of the data flow and the removal of square and root operations in batch normalization give our framework the ability to perform inference on fixed-point devices. Our proposed framework is evaluated on mainstream semantic segmentation networks like FCN-VGG16 and DeepLabv3-ResNet50, achieving comparable accuracy against floating-point framework on ADE20K dataset and PASCAL VOC 2012 dataset. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel approach that reconstructs the indoor scene in a structure-aware manner and produces two meshes with different levels of abstraction. To be precise, we start from the raw triangular mesh of indoor scene and decompose it into two parts: structure and non-structure objects. On the one hand, structure objects are defined as significant permanent parts in the indoor environment such as floors, ceilings and walls. In the proposed algorithm, structure objects are abstracted by planar primitives and assembled into a polygonal structure mesh. This step produces a compact structure-aware watertight model that decreases the complexity of original mesh by three orders of magnitude. On the other hand, non-structure objects are movable objects in the indoor environment such as furniture and interior decoration. Meshes of these objects are repaired and simplified according to their relationship with respect to structure primitives. Finally, the union of all the non-structure meshes and structure mesh comprises the scene mesh. Note that structure mesh and scene mesh preserve various levels of abstraction and can be used for different applications according to user preference. Our experiments on both LIDAR and RGBD data scanned from simple to large scale indoor scenes indicate that the proposed framework generates structure-aware results while being robust and scalable. It is also compared qualitatively and quantitatively against popular mesh approximation, floorplan generation and piecewise-planar surface reconstruction methods to demonstrate its performance. [ABSTRACT FROM AUTHOR]

• A frequency-domain segmentation algorithm to reduce the pixel deviation and enhanced sharpness is proposed. • The pixel deviation threshold is proposed to segment 3D scenes in frequency-domain along the depth. • The simulations and experiments confirm that the proposed method provides higher-quality images. The three-dimensional (3D) light-field display technology with full-parallax is an inevitable future development trend, however, still suffered from pixel deviation effect. Although depth-priority arrangement and fitting algorithm have been investigated to address it, both have caused a decrease in the resolution or sharpness, respectively. In this paper, a frequency-domain segmentation algorithm based on the pixel deviation threshold is demonstrated to reconstruct high sharpness and the resolution of 3D light-field images without pixel deviation. This algorithm segments the 3D images in references to the pixel deviation threshold and combines the backward ray-tracing algorithm and the fitting algorithm to process the various frequency-domain components. Meanwhile, pixel deviation threshold with distinct depths is proposed and analyzed, and the simulation experiments verify the validity of the pixel deviation thresholds. Simulation and experimental results demonstrate that our proposed algorithm addresses the pixel deviation issue and exhibits improved performance in other aspects compared with the current algorithm. [ABSTRACT FROM AUTHOR]

Due to uneven illumination or limitations of imaging devices, intensity inhomogeneities are more or less present in images obtained by different imaging modes. This ubiquitous intensity inhomogeneity makes image segmentation more difficult. This paper proposes a new bias field model (KPBFE) based on pre-fitting bias field estimation to deal with intensity inhomogeneity in the image segmentation. A new function for computing bias field b is proposed with K-means++ clustering algorithm. The computation method of clustering center points takes into account the average value of the grayscale within the contour of the bias field estimation and outside the contour. Meanwhile, we use a variational level set function with arctan function and a new adaptive function τ to limit the magnitude of the data driver term. Since the computation of bias field estimation is completed before the iteration and there is no convolution operation in the process, the computing speed of the proposed model is greatly increased. Experiments results show that our model can effectively segment the images with intensity inhomogeneity. Compared with some classical models, our method also has faster computation speed, higher segmentation accuracy and better initial robustness. • A new function for computing bias field is proposed. • The computation of the bias field is completed before the iteration and there is no more time-consuming convolutions in iterations. • The computing of clustering center points takes into account the average value of the grayscale within and without the contour of the bias field. • The variational level set function is optimized with an adaptive function to limit the magnitude of the data driven term. • The proposed model shows better segmentation accuracy and robustness. [ABSTRACT FROM AUTHOR]

An array of existing active contour models is prone to suffering from the deficiencies of poor anti-noise ability, initialization sensitivity, and slow convergence. In order to handle these problems, a robust hybrid active contour method based on bias correction is proposed in this research paper The energy functional is formulated through incorporating the adaptive edge indicator function and level set formulation driven by bias field correction. The adaptive edge indicator function, which is formulated based on image gradient information, is utilized to detect object boundaries and accelerate the segmentation in the homogeneous region. The level set formulation is constructed based on an improved criterion function, in which bias field information is considered. Specifically, the bias field distribution is approximated through the local mean gray value algorithm as a prior. Moreover, a new regularized function is proposed so as to maintain the stability of curve evolution. The segmentation process is implemented by the optimized energy function and the novel regularized term. Compared to previous active contour models, the modified active contour method can yield more precise, stable, and efficient segmentation results on some challenging images. • A hybrid active contour model based on bias correction is proposed. • The bias field is calculated through a new method before curve iteration. • An adaptive edge indicator function and a new regularization function is given. • Experiments show the superiority of the proposed model such as high efficiency, robustness and accuracy. [ABSTRACT FROM AUTHOR]