Zihan Peng

Image Semantic Change Detection Based on Siamese Neural Networks

• Designed a siamese U-net architecture, introduced contrastive branch and balancing parameter to improve the performance. Also designed a web page interface by Vue2 to present a demo.
• Learned the contrastive learning in depth and mastered SimCLR, MoCo, etc. Contrastive learning models, template matching, and evolutionary algorithms.
• Proposed Template Contrast for image recognition and conducted comparative effectiveness tests and potential validation. Introduced TC to the image semantic change detection task.

Muti-object tracking based on self-supervised contrastive learning.

• Introduce a double contrast learning module to CenterTrack to construct a multi-target tracking model ContrastTrack, based on self-supervised contrast learning.
• Designed Association of Feature Representation Similarity(AFRS)
• ContrastTrack defeated CenterTrack on the MOTA on MOT16, MOT17 and MOT20 dataset; especially on MOT20 dataset, which has a large number of targets, ContrastTrack had 1.2% higher MOTA than CenterTrack.
• Thesis presented at China Multimedia 2023, Kunming, China, 2023.

Multiple Classification Based on Fisher Discriminant Analysis

• Implemented Fisher discriminant analysis in MATLAB and extended it to multi-classification by using the one-versus-rest strategy.
• Learned FDA and multi-classification strategy, such as one-versus-rest, one-versus-one, and many-versus-many.
• The accuracy of binary classification reaches 86.67% on a portion of MNIST, and for multi-classification, the accuracy reaches 80.25%.