Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2022-05-11 Cooperative journals: 《计算机应用研究》
Abstract： It is difficult for traditional highlight removal algorithms to effectively deal with the processing of stepped highlight images in the stepped palletizing of factory robots. To solve this problem, based on the knowledge of conditional generative adversarial network, this paper proposes a stepped surface highlight removal network model named MSDGC-GAN (Multi-scale Spatial dense gradient cascade generative adversarial network) . In this method, the Spatial Contextual Feature Dense Block (SCFDB) aims to deeply extract the spatial background information between pixel rows and columns. In addition, the multi-scale gradient cascade structure aims to compensate for the scale feature loss in network downsampling, and this structure can endow the model with multi-scale discriminative ability while stabilizing the training gradient distribution. Based on the analysis of the classical two-color reflectance model, we apply the maximum diffuse reflectance estimation to the loss function to supervise the network training. The experimental results show that the proposed method outperforms the compared methods in both the classical highlight dataset and the self-made stepped highlight image dataset.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2020-09-28 Cooperative journals: 《计算机应用研究》
Abstract： Given a fixed graph, the graph-based clustering usually performs the eigen-decomposition to obtain the spectral eigenvectors based on which we need to conduct post-processing steps such as Kmeans or spectral rotation to obtain the final clustering assignments. This paradigm may cause two limitations: a) this two-stage strategy breaks down the connection between the graph construction and the calculation of spectral eigenvectors; and b) the l2-norm based similarity measure between spectral eigenvectors is usually sensitive to noise. To deal with these two limitations, this paper proposed a robust clustering algorithm based on joint structured graph learning and l1-norm spectral clustering, termed CLRL1. In the proposed framework, on one hand the graph learning process and the clustering process can be optimized together towards the optimum; on the other hand, the l1-norm similarity measure of spectral eigenvectors is used to improve the model robustness. Experiments on extensive benchmark data sets show the effectiveness of the proposed algorithm.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2019-05-10 Cooperative journals: 《计算机应用研究》
Abstract： The traditional way of internal attack for smartphone is easy to detected and prevented by the user. As a common audio signal, DTMF signal plays a very important role in mobile communication, but also faces severe security risk. This paper proposed an external attack method for smartphone based on DTMF signal, which could attack effectively without the user being aware and without interaction with the smartphone. Firstly, it recorded some important keystroke operation of user. Secondly, performed double-threshold endpoint detection in time domain to extract the effective area of the signal. Thirdly, converted the effective area to frequency domain by Goertzel algorithm for digital classification. Finally, all the keystroke data of the user were obtained by comparing the DTMF coding table. The experimental results show that the method can decipher more than 80% of the keystroke data under the condition of 10db signal-to-noise ratio and no interaction with the smartphone.
Subjects: Computer Science >> Integration Theory of Computer Science submitted time 2018-06-19 Cooperative journals: 《计算机应用研究》
Abstract： For image distortion classification, this paper proposes an algorithm based on Gabor wavelet and CNN (Convolutional Neural Network) . It uses the good characteristic of Gabor wavelet to extract rough feature of images firstly, and then uses the improved CNN to extract the key feature from rough feature. The main steps include: Images are preprocessed firstly (including labels setting, samples balance and samples expansion) ; Then it calculates eight directions Gabor wavelet to preprocessed images, and then add eight sub-bands to one sample for training; Finally, it uses a self-designed CNN and Softmax classifier to train the final model, and uses the methods of random gradient descent and error back propagation to optimize the parameters of convolution kernels during training. The final model is used to determine the type of image distortion, the classification accuracy on the LIVE standard image library is 95.62%, it shows that the proposed method has high accuracy and robustness.
Hits
Hits
Downloads
Comment