An incremental marine litter recognition method and system based on knowledge distillation

By using the knowledge distillation incremental learning framework, combined with the teacher-student model architecture and loss function optimization, the problems of low accuracy and insufficient real-time performance in marine debris identification technology under dynamic category changes are solved, and efficient and stable debris identification is achieved.

CN120707926BActive Publication Date: 2026-06-05WUHAN UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202510718656.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2025-05-30
Publication Date
2026-06-05
Estimated Expiration
2045-05-30

AI Technical Summary

Technical Problem

Traditional marine debris identification technologies are not accurate enough when faced with dynamically increasing types of debris and cannot meet the real-time requirements of mobile devices. Existing methods suffer from catastrophic forgetting and waste of computing resources.

Method used

We adopt an incremental learning framework based on knowledge distillation. Through a teacher-student model architecture, we use soft objectives and soft predictions to generate loss functions and combine gradient descent to optimize student model parameters, thereby achieving knowledge transfer and updating and avoiding full retraining.

Benefits of technology

It significantly improves the model's detection accuracy and stability for new garbage categories, reduces computational resource consumption and training time, and is suitable for resource-constrained mobile platforms.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120707926B_ABST
    Figure CN120707926B_ABST
Patent Text Reader

Abstract

The application discloses an incremental marine garbage recognition method and system based on knowledge distillation, and belongs to the field of deep learning, and comprises the following steps: acquiring a marine garbage data set to be recognized; inputting the marine garbage data set to be recognized into a trained student model to output a marine garbage recognition result; wherein the training of the student model comprises the following steps: inputting an updated data set into a teacher model to generate a soft target and a new soft target; initializing the student model by using the teacher model, inputting the updated data set into the student model to generate a soft prediction and a new soft prediction; calculating a hard label loss from a real label and the new soft prediction, calculating a distillation loss based on the soft target, the soft prediction, the new soft target and the new soft prediction, updating parameters of the student model based on the distillation loss and the hard label loss, and stopping until the student model converges. The application not only solves the problems of excessive redundant information and catastrophic forgetting of the model, but also improves the freedom degree of model recognition.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Knowledge distillation-based edge device scene identification method and device

    CN114241282A

  • Target detection incremental learning method based on high-quality pseudo label mining

    CN118968165A