Search

Soybeans have the characteristics of balanced amino acid species and high nutritional value. In this article, the feasibility of the identification soybean from three typical origins (Argentina, the United States and China) by interval partial least squares (iPLS) optimized terahertz (THz) spectroscopy combined with chemometrics was investigated. Firstly, the THz frequency-domain spectrum was optimized using iPLS. Then, 168 soybean samples were selected as the correction set, and soybean origin identification models were respectively built using the extreme learning machine (ELM), genetic algorithm support vector machine (GA-SVM), and artificial bee colony algorithm support vector machine (ABC-SVM) combined with 8 pre-processing techniques. Finally, the models were verified through 57 samples of the test set and the comprehensive identification accuracy rate of the ABC-SVM model reached 94.74%. The experimental results showed that after iPLS optimization and appropriate pre-processing technique, THz spectroscopy and chemometrics could accurately identify the origin of soybean.

The stereo matching technology has been widely used nowadays for robot guidance, depth of field rendering, and video processing applications. The binocular stereo matching technology collects two images of the same scene from different views, computes the disparity, and then determines the 3-D depth of the object using the triangulation theory. In this paper, we propose a new deep-learning-based stereo matching algorithm called convolutional neural network WTADP (CNN-WTADP). The commonly used WTA method ignores the disparity constraint of the neighboring pixels and is prone to be affected by noise, so the obtained original disparity map may contain a lot of abnormal values and mismatched points. To solve this problem, we construct a CNN for matching cost computation and use a novel global method of dynamic programming based on WTA for disparity selection. Steps for the CNN-based computation of the matching cost are presented, and the post-processing algorithm is used to obtain the final determination of the disparity map. The Middlebury stereo dataset and stereo evaluation are used to compare the proposed algorithm with other methods. The experiments and its results demonstrate the ability of the proposed algorithm to match the images more effectively.

Depth image-based rendering (DIBR) is one of the main techniques in interactive 3D video systems. But a critical problem of DIBR is that how to eliminate the holes appearing in virtual view images. In this paper, an improved depth image based virtual view synthesis algorithm is proposed. In depth map pre-processing, only local areas of the depth map with sudden changes are pre-processed by the asymmetric filter. After this, we combine the forward and inverse 3D-warping to synthesize the virtual view. Firstly, it warps the left image to the virtual view and then only reversely warps the region of the holes in the virtual view to the right view to get the lost information of holes. Compared with the traditional method which have to warp the two whole images, computation cost can be reduced by 30% efficiently. Finally, we employ an improved image inpainting based on depth information to fill hole regions. We have improved the method in two portions: 1) depth information is added to the priority computation and searching the best matching patch and 2) as the filling proceeds, the confidence term of the newly inpainted pixels is updated with the mean square error (MSE) of pixels in the target matching patch and the source matching patch. Experimental results show significant objective and subjective gains of the proposed method in comparison to the reference methods.

Infrared image segmentation is a useful and challenging research subject due to its inherent limitations, such as complicated noises, vague edges, and low contrast in infrared images. Active contour models have a wide range of applications for infrared image segmentation. A gradient vector flow (GVF) model and other external force field models have been proposed to improve the noise robustness and weak edge protection to a certain extent. To further address these issues, this paper presents a novel edge-preserving active contour model using guided filter and harmonic surface function for infrared image segmentation. Guided filter is applied to obtain the edge map, reducing the noise interference effectively and collecting more detailed information of the infrared images such as the edges. Then, a harmonic surface function is added into the smoothness term to make the proposed model possess both the abilities of fast convergence and weak edge protection. Besides, we incorporate the information of external force field into the proposed model to preserve the weak boundaries. Compared with adaptive diffusion flow, the proposed method has increased 4% and 2% in precision and F1 measure, respectively.

At present, a hyperspectral image has a significant advantage in the aspect of application because of its high spectral resolution. However, owing to the limit of communication capacity, a hyperspectral image must be compressed. In this paper, we develop and propose a fractal lossy hyperspectral image compression method based on prediction. First, we exploit spatial correlation by applying predictive lossy compression to obtain a reference band of high quality. Then, the local similarity between the two adjacent bands is used through fractal encoding using a local search algorithm. Next, we encode the fractal parameters and the error and fractal residual is transformed by discrete cosine transform, quantized, and entropy encoded to improve the decoded quality. Through experiments, we demonstrate that the proposed algorithm leads to considerably improved performance in compression compared with the other well-known methods. Finally, we validate whether the compression affects the data in the hyperspectral images through classification. The results indicate that the accuracy of classification obtained for the reconstructed image is marginally less than the accuracy reported for the original data set; however, the loss in accuracy is less than 1% and thus acceptable.

Video compression has become more and more important nowadays along with the increasing application of video sequences and rapidly growing resolution of them. H.264 is a widely applied video coding standard for academic and commercial purposes. And fractal theory is one of the most active branches in modern mathematics, which has shown a great potential in compression. In this paper, this study proposes an improved inter prediction algorithm for video coding based on fractal theory and H.264. This study take the same approach to make intra predictions as H.264 and this study adopt the fractal theory to make inter predictions. Some improvements are introduced in this algorithm. First, luminance and chrominance components are coded separately and the partitions are no longer associated as in H.264. Second, the partition mode for chrominance components has been changed and the block size now rages from 16 × 16 to 4 × 4, which is the same as luminance components. Third, this study introduced adaptive quantization parameter offset, changing the offset for every frame in the quantization process to acquire better reconstructed image. Comparison between the improved algorithm, the original fractal compress algorithm and JM19.0 (The latest H.264/AVC reference software) confirms a slightly increase in Peak Signal-to-Noise Ratio, a significant decrease in bitrate while the time consumed for compression remains less than 60% of that using JM19.0.