Aiming at the limitations of traditional noise reduction methods for acoustic signals in low Signal-to-Noise Ratio (SNR) environment, a real-time noise reduction algorithm was proposed by combining adaptive threshold Voice Activity Detection (VAD) algorithm and Minimum Mean-Square Error Log-Spectral Amplitude estimation (MMSE-LSA). Firstly, the background noise was estimated in VAD algorithm by probability statistics based on the maximum value of the energy probability, and the obtained background noise was updated in real time and saved. Then, the background noise updated in real time was used as the reference noise of MMSE-LSA, and the noise amplitude spectrum was updated adaptively. Finally, the noise reduction processing was performed. The experimental results on four kinds of acoustic signals in real scenes show that the proposed algorithm can guarantee the real-time processing of low SNR acoustic signals; and compared with the traditional MMSE-LSA algorithm, it has the SNR of the noise reduction signal increased by 10-15 dB without over-subtraction. It can be applied to practical engineering.

Aiming at the problems of insufficient generalization ability, poor stability and difficulty in meeting the real-time requirement of facial expression recognition, a real-time facial expression recognition method based on multi-scale kernel feature convolutional neural network was proposed. Firstly, an improved MSSD (MobileNet+Single Shot multiBox Detector) lightweight face detection network was proposed, and the detected face coordinates information was tracked by Kernel Correlation Filter (KCF) model to improve the detection speed and stability. Then, three linear bottlenecks of three different scale convolution kernels were used to form three branches. The multi-scale kernel convolution unit was formed by the feature fusion of channel combination, and the diversity feature was used to improve the accuracy of expression recognition. Finally, in order to improve the generalization ability of the model and prevent over-fitting, different linear transformation methods were used for data enhancement to augment the dataset, and the model trained on the FER-2013 facial expression dataset was migrated to the small sample CK+ dataset for retraining. The experimental results show that the recognition rate of the proposed method on the FER-2013 dataset reaches 73.0%, which is 1.8% higher than that of the Kaggle Expression Recognition Challenge champion, and the recognition rate of the proposed method on the CK+ dataset reaches 99.5%. For 640×480 video, the face detection speed of the proposed method reaches 158 frames per second, which is 6.3 times of that of the mainstream face detection network MTCNN (MultiTask Cascaded Convolutional Neural Network). At the same time, the overall speed of face detection and expression recognition of the proposed method reaches 78 frames per second. It can be seen that the proposed method can achieve fast and accurate facial expression recognition.

In order to quickly and stably detect the faces with large change of angle and serious occlusion in image sequence, a new automatic Detection-Tracking-Detection (DTD) model was proposed by combining the fast and accurate target detection model MobileNet-SSD (MS) and the fast tracking model Kernel Correlation Filtering (KCF), namely MS-KCF face detection model. Firstly, the face was detected quickly and accurately by using MS model, and the tracking model was updated. Secondly, the detected face coordinate information was input into the KCF tracking model to track steadily, and the overall detection speed was accelerated. Finally, to prevent tracking loss, the detection model was updated again after tracking several frames, then the face was detected again. The recall of MS-KCF model in the FDDB face detection benchmark was 93.60%; the recall in Easy, Medium and Hard data sets of WIDER FACE benchmark were 93.11%, 92.18% and 82.97%, respectively; the average speed was 193 frames per second. Experimental results show that the MS-KCF model is stable and fast, which has a good detection effect on the faces with serious shadows and large angle changes.

As an important component of the surface meteorological observation, the daily observation of surface frost still relies on manual labor. Therefore, a new method for detecting frost based on computer vision was proposed. First, a k-nearest neighbor graph model was constructed by incorporating the manually labeled frosty image samples and the test samples which were acquired during the real-time detection. Second, the candidate frosty regions were extracted by rating those test samples using a graph-based manifold learning procedure which took the aforementioned frosty samples as the query nodes. Finally, those candidate frosty regions were identified by an on-line trained classifier based on Support Vector Machine (SVM). Some experiments were conducted in a standardized weather station and the manual observation was taken as the baseline. The experimental results demonstrate that the proposed method achieves an accuracy of 87% in frost detection and has a potential applicability in the operational surface observation.