The performance of salient object detection is greatly improved by the superior feature extraction ability of Fully Convolutional Neural Networks(FCN).However，the simple fusion strategies (feature addition or concatenation) cannot effectively enhance features，resulting in algorithms object misdetection and missed detection in complex scenes.The paper proposed a specifically feature enhancement method to improve the performance of salient object detection.Firstly，object misdetection mostly occurs in a scene where the background is cluttered or the object and the background are intertwined，so it greatly alleviate the object misdetection problem from the perspective of global enhancement and structural enhancement，respectively.Secondly，the missed detection of the object generally occurs in the interior and edge of the object，so the study introduce residual learning to learn the information of the missed region and refine the loss of the object interior and edge.Finally，comparison results between the proposed method with other 13 kinds of advanced methods over 5 benchmark datasets indicate that the proposed model is superior to other 13 methods，and the problems of object misdetection and missed detection in complex scenes were successfully solved.

Based on the fact that most steganography approaches follow the rules of minimum distortion or maximum entropy，this paper proposed to employ the distribution of optimal embedding probability as the prior knowledge for video steganalysis.To better characterize the embedding priority of motion vectors，it defined a measurement of the embedding distortion of motion vectors using features from three aspects，namely motion feature，texture feature and local optimality under coding framework，and the embedding probabilities of motion vectors were estimated with Gibbs distribution.Thus this study proposed a way of quantitatively enhancing the steganalytic features with the embedding probabilities and the mechanism of the enhancement was explained from the perspective of relative entropy.Experimental results show that the detection accuracy of three classical steganalytic methods have been unanimously improved and show robustness against different bitrates after the enhancement with the proposed method.The effectiveness of the new method is also verified by the comparison with the latest deep neural network VSRNet detection method.

Tactile sensors have received extensive attention and research because they can imitate human skin to perceive external information.Among varies kinds of tactile sensor，Nano-ZnO-based tactile sensor has broad application prospects in the fields of bionic robots，human-computer interaction，biomedicine devices and wearable electronic systems due to its advantages of high sensitivity，high mechanical strength，strong flexibility，easy integration，good biocompatibility，and low-cost technology.This paper firstly introduced the piezoelectric mechanism and the manufacturing of Nano-ZnO-based tactile sensor.Next，it summarized the study progress of Nano-ZnO piezoelectric tactile sensor from the four aspects of piezoelectric performance improvement，multi-function，integration，and array fabrication，and listed the latest achievements in related technical fields and research directions.Then it summarized the application research of Nano-ZnO piezoelectric tactile sensor in pressure imaging，human action detection and physiological health monitoring in detail，and analyzed the shortcomings and the research trend of Nano-ZnO piezoelectric tactile sensor.

Music source separation is one of the most important research topics in the field of music information retrieval.Traditional music source separation methods have shortcomings，such as hypothesis dependence，limited model complexity，and poor representation ability.To resolve these problems，it takes a long time to train the time-domain end-to-end deep learning network model，and the separation performance still needs to be improved.Therefore，in order to further optimize the representation ability and computational efficiency of the time domain end-to-end separation model，the study proposed an end-to-end network Unet-SE-BiSRU based on the Demucs model which has the best performance in time domain separation at present.Attention mechanism was introduced into the generalized coding layer and decoding layer，and the squeezing-excitation block(SE) was used to extract features selectively according to the type of audio to be separated.To deal with gradient explosion or disappearance that may occur in the learning process，a group normalization was added after one-dimensional con-volution.The bidirectional long short-term memory network was refined to a bidirectional simple recurrent unit(BiSRU)，which improves the parallelism of learning and reduces the amount of model parameters.The experimental results show that the signal-noise ratio of the improved network model is improved by 0.34dB，which is the best one among the time-domain end-to end methods to the best of our knowledge，and the training time is reduced by 3/5.

The development of 5G has promoted the miniaturization and integration of communication equipment，but it also leads to new electromagnetic compatibility problems.Introducing Frequency Selective Surface (FSS) into devices can effectively reduce the interference between devices.In order to overcome the weaknesses of traditional design methods，such as low efficiency，poor quality and dependence on experience，a new FSS shape optimization method was proposed herein.The proposed method presented a strategy to define the shape optimization parameters in the polar coordinate system according to the non-self-intersecting characteristics of FSS.Furthermore，based on the genetic algorithm，the method realized the intelligent search of parameters optimized by FSS and realized the intelligent optimazition of FSS unit shape in terms of the operating frequency and electromagnetic shielding performance.Three numerical examples including solid，ring and slot units have been used to verify the proposed method.The results show that the optimized FSS resonant frequency moves to 28GHz and its electromagnetic shielding performance is improved.The FSS obtained by the proposed method is stable and suitable for 5G environment.

The existing deep learning-based speech bandwidth extension algorithms have many disadvantages：the time domain algorithms speech  feature extraction  is not accurate enough and its training data is too large；the frequency domain algorithm pays little attention to the information association between frames in log power spectrum feature extraction and the number of frequency axes is odd number which is inconvenient for deepening the network depth.In addition，it ignores time domain information；the time-frequency two-domain algorithm model is relatively complicated.To solve these problems，this paper proposed a speech bandwidth extension algorithm based on Flatten-CNN.Firstly，in order to make full use of speech features and reduce the amount of data，the algorithm was operated on frequency domain.Secondly，an improved encoder was proposed  to make use of the logarithmic power spectrum time axis information.The log power spectrum feature extraction of two-axis was realized by introducing tile layers.Thirdly，in order to deepen the network depth，the last point was removed during the frequency axis data processing and a zero was added when restoring，so  to ensure that the frequency axis number is an even number.Finally，in order to utilize the voice signal time domain information，time domain loss was introduced into the loss function.The effectiveness of the algorithm  was verified with the TIMIT data set and the VCTK data set.The experimental results show that，compared with the current mainstream algorithms，the new algorithm can improve the high-bandwidth speech quality，showing better hearing effect.

In the infrared gas detection system based on tunable diode laser absorption spectroscopy (TDLAS), the second harmonic signal is used to calculate the gas percentage by volume. However, this signal is easily influenced by many parameters of the hardware. Based on Lambert Beers law and HITRAN database, this paper used LabVIEW to study the influence of parameters on the amplitude, symmetry and peak width of the second harmonic, and carried out the harmonic optimization simulation experiment. The results show that, for the general TDLAS detection system, the appropriate scanning amplitude and filter order should be chosen firstly, so that the harmonic line would be complete and undistorted. Furthermore, the amplitude and peak width of the harmonic should be increased as much as possible to get the optimal modulation depth. Finally, the suitable modulation and scan frequency should be determined choose according to the system requirements and the hardware conditions.

The current active probe, whose detection frequency is mainly concentrated in the low frequency band, cant meet the detection requirements of the high frequency band. Therefore, a small, high-bandwidth, non-contact active magnetic field probe was proposed. The active magnetic field probe is made of a multilayer printed circuit board (PCB). An active amplifier module and its supporting power management chip were added to the passive probe to improve the transmission gain of the ultra-wideband type probe. The probe was tested and analyzed from four aspects: frequency response, spatial resolution, calibration factor, and differential electric field suppression capability. The results show that the designed probe has a transmission gain of -20dB, a spatial resolution of 900μm, and is of good differential electric field suppression. The probe can be used for ultra-wideband PCB board and more complex integrated circuit measurements.

Facing the high data volume and low latency requirements of mobile Virtual Reality(VR) services, this paper established a partial cache content sharing model for mobile VR and simulated the social attribute-based content sharing cost minimization problem, in order to effectively reduce the data traffic of the core network and reduce the requesting cost of devices. The problem was decomposed into two sub-problems, namely, power optimization with mode selection and device association. A cost-based power optimization and mode selection algorithm and a matching algorithm based on social relationship and requesting cost were respectively proposed. Finally, the performance of proposed algorithms was evaluated through simulations and was compared with other schemes. The simu-lation results show that the content request mode selection and power optimization scheme based on partial cache can effectively reduce the requesting cost of content requesters.

Edge-preserving image smoothing is a key preprocessing step of many computer vision and graphics algorithms. Edge-preserving image smoothing algorithm based on l1 norm outperforms many existing image smoothing algorithms. However, there are still many remaining textures in the smoothed results. In order to improve the smoothing effect of the algorithm, a new image smoothing algorithm based on reweighted l1 norm was proposed. The proposed algorithm combines the original l1 norm-based image smoothing algorithm and reweighed l1 norm minimization. The solution was made sparser through the reweighting method. The performance of the algorithm was evaluated through experiments. Experimental results show that, compared to the original l1 norm-based image smoothing algorithm and several other existing image smoothing algorithms, the proposed algorithm can smooth images very effectively and there is little texture remained in the smoothed results. Using reweighted l1 norm can improve the smoothing effect of the original l1 norm-based image smoothing algorithm.

In the wastewater treatment process, due to the complexity of the industry process, incomplete monitoring equipment and hostile working environment, it is difficult to achieve accurate and timely measurement of important effluent indices. To solve the problem, an ensemble adaptive soft sensor model based on semi-supervision learning was proposed. Firstly, Gaussian process regression (GPR) and Relevance vector machine (RVM) were used to establish a heterogeneous soft-sensor model. Then, the structure and parameters of the models  were optimized by using the moving window and the Kalman filter gain, respectively. Finally, the prediction performance and self-adaptability of the model were verified by experiments on a wastewater plant. The results demonstrate that the proposed method improves the prediction accuracy and self-adaptability of the soft sensor model.

Traditional Compressed Video Sensing (CVS) reconstruction algorithm is highly time-consuming. Newly developed CVS neural networks can successfully deal with the speed problem, but it fails to make full use of the spatiotemporal correlation of video and leads to a poor performance. To solve this problem, a novel two-stage multi-hypothesis neural network (2sMHNet) was proposed. Firstly, the Temporal Deformable Alignment Network（TDAN）was used to realize pixel based multi-hypothesis prediction. While avoiding block effects, it improves the matching accuracy of the hypothesis set and obtains accurate multi-hypothesis weights by adaptively parameters learning. Then, the residual reconstruction module was constructed to reconstruct the prediction residual with measurements to further improve the reconstruction quality. Finally, in order to make full use of the inter-frame correlation, a two-stage serial reconstruction mode was proposed. In the first stage, as the reconstructed key frames have rich details, they are selected as the reference frame to improve the non-key frames quality. In the second stage, the more relevant adjacent frames are used for motion compensation, which is more conducive to fast and complex sequences. Experimental results demonstrate that the proposed 2sMHNet outperforms the existing good CVS reconstruction algorithms.