We present a real-time algorithm for foreground–background segmentation. Sample background values at each pixel are quantized into codebooks which represent a compressed form of background model for a long image sequence. This allows us to capture structural background variation due to periodic-like motion over a long period of time under limited memory. The codebook representation is efficient in memory and speed compared with other background modeling techniques. Our method can handle scenes containing moving backgrounds or illumination variations, and it achieves robust detection for different types of videos. We compared our method with other multimode modeling techniques. In addition to the basic algorithm, two features improving the algorithm are presented—layered modeling/detection and adaptive codebook updating. For performance evaluation, we have applied perturbation detection rate analysis to four background subtraction algorithms and two videos of different types of scenes

Automatic liver tumor segmentation is a paramount important application for liver tumor diagnosis and treatment planning. However, it has become a highly challenging task due to the heterogeneity of the tumor shape and intensity variation. Automatic liver tumor segmentation is capable to establish the diagnostic standard to provide relevant radiological information to all levels of expertise. Recently, deep convolutional neural networks have demonstrated superiority in feature extraction and learning in medical image segmentation. However, multi-layer dense feature stacks make the model quite inconsistent in imitating visual attention and awareness of radiological expertise for tumor recognition and segmentation task. To bridge that visual attention capability, attention mechanisms have developed for better feature selection. In this paper, we propose a novel network named Multi Attention Network (MANet) as a fusion of attention mechanisms to learn highlighting important features while suppressing irrelevant features for the tumor segmentation task. The proposed deep learning network has followed U-Net as the basic architecture. Moreover, residual mechanism is implemented in the encoder. Convolutional block attention module has split into channel attention and spatial attention modules to implement in encoder and decoder of the proposed architecture. The attention mechanism in Attention U-Net is integrated to extract low-level features to combine with high-level ones. The developed deep learning architecture is trained and evaluated on the publicly available MICCAI 2017 Liver Tumor Segmentation dataset and 3DIRCADb dataset under various evaluation metrics. MANet demonstrated promising results compared to state-of-the-art methods with comparatively small parameter overhead.

The statistical background subtraction and shadow detection algorithm (SBGS) is fast and reliable in outdoor scenes with shadows. However, its reliability depends on the number of training frames to construct the initial background model. In addition, the similarity between foreground and background colors, i.e, camouflage problem, could lead to the worse performance of background subtraction. In this paper, we present a robust outdoor background subtraction technique based on color statistics and edge information. Vector median filtering technique was employed in the initialization of the background model to address the SBGS's limitation. In addition, a combination of color statistics and edge information is utilized to improve the segmentation results over the original algorithm. Test data was compiled from various outdoor conditions including strong shadow, complex background, and low contrast scenes. The background subtraction results show that the proposed approach outperformed other well-known segmentation algorithms such as non-adaptive and adaptive SBGS algorithms as well as mixture of Gaussian algorithm based on precision-recall and computational measurements.

This paper describes a vision system that can extract 2D and 3D visual properties of mango such as size (length, width, and thickness), projected area, volume, and surface area from images and use them in sorting. The 2D/3D visual properties are extracted from multiple view images of mango. The images are first segmented to extract the silhouette regions of mango. The 2D visual properties are then measured from the top view silhouette as explained by Yimyam et al. (2005). The 3D mango volume reconstruction is done using volumetric caving on multiple silhouette images. First the cameras are calibrated to obtain the intrinsic and extrinsic camera parameters. Then the 3D volume voxels are crafted based on silhouette images of the fruit in multiple views. After craving all silhouettes, we obtain the coarse 3D shape of the fruit and then we can compute the volume and surface area. We then use these features in automatic mango sorting which we employ a typical backpropagation neural networks. In this research, we employed the system to evaluate visual properties of a mango cultivar called "Nam Dokmai". There were two sets total of 182 mangoes in three various sizes sorted by weights according to a standard sorting metric for mango export. Two experiments were performed. One is for showing the accuracy of our vision-based feature extraction and measurement by comparing results with the measurements using various instruments. The second experiment is to show the sorting accuracy by comparing to human sorting. The results show the technique could be a good alternative and more feasible method for sorting mango comparing to human's manual sorting.

This paper describes our algorithms for players tracking and ball detection for an automatic broadcast tennis video annotation. The system detects and tracks the players using a robust non-parametric procedure for estimating density gradients called the mean shift algorithm. The basic mean shift tracking algorithm assumes that the target object has to separate sufficiently from background, but this assumption is not always true especially when tracking is carried out in dynamic backgrounds such as in sport videos. To cope with this problem, in our proposed system, we embrace the motion segmentation and use the 8×8×8 color histogram to be feature distribution for mean shift tennis players tracking. In order to determine the players' actions precisely, the system also detect and track ball positions using frame differencing as well as applying some correlation techniques to eliminate false detections. Based on both players' motion patterns and ball positions, the system can precisely classify the players' action into backhand ground stroke and forehand ground stroke. Videos of broadcast tennis games downloaded from the Internet have been tested. The results show our system is able to precisely classify the player's actions with 83.7% precision and 82% recall rates.