Search

Rotating orthogonal polarization imaging (ROPI) is an effective de-scattering method to detect objects through turbid media. This paper investigates improving the performance of ROPI. On the basis of the Jaffe model, polarization filtering using the Stokes-Mueller formula is performed theoretically and demonstrated experimentally. The results show that the illumination polarization angle and difference in the Mueller matrix elements of m12, m13, m21, and m31 between the medium and the object could affect the signal-to-noise ratio of images obtained via the ROPI method. In addition, polarization filtering by rotating the polarizer mechanically could be replaced by Mueller matrix based polarization information processing for the precise control of the illumination polarization angle and rapid imaging process.

Temporal action detection is used to detect the start and end times and classify the potentially specific actions in a video. Prior studies in temporal action detection perform weak because they can not fully understand the whole input video's temporal structure and context information, and fail to adapt to the diversity of action time span. We propose a novel Timeception Single Shot Action Detector (TC-SSAD) to solve the problems mentioned above. In detail, we leverage the multiple Timeception layers to generate multi-scale feature sequences, where each Timeception layer uses depthwise-separable temporal convolution with multi-scale convolution kernels to capture the diversity of time spans. Besides, we use the super-event modules to learn the entire input video’s temporal structure and contextual information. The experimental results on THUMOS14 dataset show that when IoU threshold is 0.5, our method achieves 38.2% and 44.3% mAP on Two-stream features and Two-stream i3D features respectively, which is better than Decouple-SSAD network based method by 2.4% and 0.6%. Our method on Activitynet-1.3 dataset achieves 20.4% mAP, which is better than Decouple-SSAD network based method by 0.61% as far as Two-stream features on concerned.

Conventional facial expression recognition methods usually deal with frontal face images via only one or several features, which are easy to loss useful information and sensitive to face poses, scales and noise. As an interesting application of facial expression, this paper proposes an effective smile detection method under unconstrained scenarios, employing convolution neural network to learn and automatically extract discriminative features from a large number of human face images. Specifically, our method firstly converts the original color images to grayscale images, and due to the important role of mouth in expression analysis, we then localizes the mouth region according to 5 key points on the face. After the brightness adjustment and size normalization, the mouth images are input as training images of an improved LeNet-5 model to learn and automatically extract the discriminative features of the mouth regions. Finally, a SVM classifier is trained to distinguish smiling or non-smiling. Experimental results of the public MTFL database and GENKI-4K database show that the accuracy rates of our method are up to 87.81% and 86.80%, respectively.

As a kind of multiplicative noise, Speckle often degrades image quality. To estimate speckle noise helps to image restoration and consequent analysis. This paper suggests a Gaussian-Hermite moments based method to evaluate the variance of speckle. In the method, the characteristics of Gaussian-Hermite moments are discussed first. Then speckle noise with different variances are respectively added to a certain image where all the greyscales of pixels are the same, and the distribution of feature vectors based on Gaussian-Hermite moments is employed to analyze the noise intensity. Next, a conception of noise characteristic value is used to stand for noise intensity and herefrom construct a function mapping of noise variances and noise characteristic values, with which an estimate function is finally established via a polynomial curve fitting. Experimental results indicate that the estimation function can rapidly and correctly provide with the intensity of speckle without any prior knowledge. Potential applications include noise assessment, image quality analysis, and guidance for noise reduction.

Three subjective evaluation experiments involving in target recognition, semantic differential (SD) and adaptive grouped paired comparison (AGPC) were designed and performed to investigate the effect of timbre attributes on acoustic target discrimination. Firstly, 50 sound samples of equalized loudness and duration were used to subjectively identify three kinds of targets. Then SD and AGPC based target discrimination experiments were carried out, and the experimental data were processed using coherence and clustering approach to explore the roles of the two kinds of timbres, natural timbre and essential timbre, in the identification of acoustic targets. This work provides some insight into timbre mechanisms inherent in auditory object identification and timbre modeling applied to auditory feature extraction and target recognition.

The influence of noises is obviously in the image transform method based on the pixels elements representation of images. To some extent the method of image transform based on linear elements representation can solution the problem. The Beamlet transform is an effective method of line segment extraction. This work improved the traditional Beamlet transform by considering the directional information of lines. The improved method transformed a digital image to a coefficient matrix. This method can embody the linear singularity of some linear targets and can be used for edge detection and extracting other useful targets in noisy images. The experimental results on manual images and SAR images demonstrate the effectiveness of this method.

This paper presents a new method for segmentation of synthetic aperture radar (SAR) images. Based on a non-parametric Bayesian infinite mixture model, Dirichlet process mixture model cluster method is proposed to segment SAR image. The traditional finite mixture model segmentation method is adapted extensively in SAR image segmentation, but the performance and the robustness is not good enough. However, the proposed infinite mixture model can simulate the intrinsic property of SAR image and the segmentation method can determine the cluster number automatically. The experiment results on the simulated data and real data show that the proposed method gets comparative performance and robustness with the traditional methods.

Speckle noise appears in synthetic aperture radar (SAR) images owing to the SAR imaging mechanism. This paper investigates and proposes a novel method on SAR images despeckling via grey system theory. In the method, we dynamically select one referential sequence to stand for inner region pixels, and a group of comparative sequences to represent the pixels to be enhanced. Then, edge pixels are distinguished from non-edge pixels via the grey relational degrees between the two kinds of sequences, and kept unchanged; while the noise and inner region pixels, taken as non-edge pixels, are adjusted to some new values. Experimental results show that the method, when being applied to both simulated and real SAR images, has a good performance in peak signal-to-noise ratio (PSNR) improvement, and outperforms most of the conventional filters: mean filter, median filter, Lee filter, Kuan filter and Frost filter.