Yingchun Liu

Aiming at the problems that the existing networks do not adapt to the image characteristics of edge scenes, do not effectively process the detail information in images and have insufficient feature fusion, this thesis proposes a feature extraction network model SFPN that can better utilize the different depths and detail semantic information of the network. Firstly, the S-soft unit is designed to process the detail information of the feature map. Secondly, this thesis proposes a new secondary fusion strategy that combines the features of the middle layer are fused with the lower and higher layers respectively, and with this, SFPN Block is proposed as an auxiliary classifier, which participates in the final prediction by weighting, fully considering the prediction information of different layers of the network. Finally, a joint loss function is introduced for training, so that the feature extraction ability of the network for targets with large scale changes and small sizes is improved, and it is more adapted to the edge computing scenario of detection requirements and improve the accuracy of feature extraction.

In this paper, we proposed an automated speech recognition system focus on the dispatching call recordings in the underground coal mines which promoted the development of intelligent coal mining. The main challenges of the speech recognition system are the noise of recordings and the dialect speech. We employed a voice activity detection module to preprocess the recordings, this module is able to reduce the noise and segment the long recording speech; then the Conformer model with CTC algorithm is utilized to train the ASR module. To get better performance, the WenetSpeech pretrained model is embedded for fine-tuning. The result shows that compared with the other general speech recognition systems, our ASR system has great advance in recognizing the dispatching call recordings of Huaibei dialect.

Aiming at the problem of low detection and recognition rate caused by crowded occlusion or indistinct difference between front and back backgrounds when performing human pose estimation in crowded scenes, We proposed an improved YOLO-Pose algorithm that combines convolutional block attention modules. By Integrate CBAM to Bottleneck module to dig deep into the feature information of the two levels of channel and space, and enhance the feature extraction of the target object in the image. In order to improve the convergence ability of the model, improve the CIOU loss function, use EIOU instead of the original regression function of YOLOv8, optimize the regression accuracy of anchor points, reduce the difficulty of network training, and improve the detection rate of occlusion. Compared with the general YOLOv8 algorithm paper, the improved algorithm proposed in this paper has a higher accuracy rate when performing pose estimation in crowded scenes.

Human pose estimation is an important research direction in the field of computer vision, which is widely used in human-computer interaction, behavior analysis, and intelligent surveillance. Although existing human pose estimation algorithms possess high accuracy for high-resolution images, they perform poorly for low-resolution images that are prevalent in practical applications, and thus are difficult to be widely used in people's daily lives. In this paper, we propose a new end-to-end network framework for accurate human pose estimation of low-resolution images by combining super-resolution assistance and quantization error optimization. In addition, a composite loss function is designed to jointly train the super-resolution network to generate high-resolution images that contribute to human pose estimation instead of simple pre-processing. The experimental results show that the mAP of our method reaches 68.1% and 61.4% on the COCO datasets downsampled to 128×96 and 64×48.