Jun Su

Mx. Jun Su


Research Focus & Metrics

Academic Career Length

4Years (2023-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
15 November 2023 in Sensors
Verified
Journal Article
Open access
PDF file
JSJun Su
Jun Su
WHWei He
Wei He
YWYingguan Wang
Yingguan Wang
RMRunze Ma
Runze Ma
One-shot object detection has been a highly demanded yet challenging task since the early age of convolutional neural networks (CNNs). For some newly started projects, a handy network that can learn the target’s pattern using a single picture and automatically decide its architecture is needed. To specifically address a scenario in which a single or multiple targets are standing in relatively stable circumstances with hardly any training data, where the rough location of the target is required, we propose a one-shot simple target detection model that focuses on two main tasks: (1) deciding if the target is in the testing image, and (2) if yes, outputting the target’s location in the image. This model requires no pre-training and decides its architecture automatically; therefore, it could be applied to a newly started target detection project with unconventionally simple targets and few training examples. We also propose an architecture with a non-training parameter-gaining strategy and correlation coefficient-based feedforward and activation functions, as well as easy interpretability, which might provide a perspective on studies in neural networks. We tested this design on the data we collected in our project, the Brown–Yosemite dataset and part of the Mnist dataset. It successfully returned the target area in our project and obtained an IOU of up to 87.04%, reached 80.28% accuracy on the Brown–Yosemite dataset with disposable networks, and obtained an accuracy of up to 89.4% on part of the Mnist dataset in the detection task.

21 October 2024 in Sensors
Verified
Journal Article
Open access
PDF file
JSJun Su
Jun Su
WHWei He
Wei He
YWYingguan Wang
Yingguan Wang
ZBZhiyong Bu
Zhiyong Bu
TZTiantian Zhang
Tiantian Zhang
In research on building a one-shot learning neural network without pre-training using mass data, the limitation on the information obtained from a single training sample downgrades the performance of the network. In order to improve performance and take full advantage of the support set, in this study, we design three kinds of shadow nodes and propose a structure-based training method for a correlation-coefficient-based neural network. This training strategy focuses on branches that are not activated or inactivated as expected. In contrast to existing networks that optimize the parameters using back-propagation, the training method proposed in this paper optimizes the structure of the correlation-coefficient-based network by correcting its pixel errors. For the shadow nodes and training process based on this strategy, the intersection over union (IOU) of a detected target increases by 4.83% in the experiments when using the Fashion-Mnist dataset, increases by 4.02% when using the Omniglot dataset, and increases by 3.89% when using the Cifar-10 dataset. The samples in category “7” wrongly classified as “1” decreased by 27.32% when using the Mnist dataset after training. This training strategy, along with shadow nodes, makes the correlation-coefficient-based network a more practical model and enables the network to develop during the accumulation of reliable samples, thus making it more suitable for simple target detection projects that collect samples over time. Moreover, the shadow nodes and training method proposed in this paper supplement the non-gradient-based parameter-gaining strategy. Additionally, it is a new attempt to explore the imitation of a human’s ability to learn a new pattern from a low number of references.