Structure grouping de-correlation pruning method and infrared small target lightweight detection method

CN120806025AActive Publication Date: 2025-10-17NAT UNIV OF DEFENSE TECH
3 Cites -1 Cited by

Patent Information

Application Number
CN202511309328.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-09-15
Publication Date
2025-10-17
Estimated Expiration
2045-09-15

AI Technical Summary

Technical Problem

Existing pruning methods have failed to effectively address the characteristics of infrared small target detection tasks, leading to the accidental deletion of key structures or a decline in model performance, and they are difficult to adapt to parameter sharing mechanisms in cross-layer connections.

Method used

By setting up a shielding layer to protect the key structure of the model, the model identifies and forms strongly correlated substructures based on the coupling dependency relationship of neurons, groups are divided according to structural similarity, and filter weights and bias terms are optimized by combining decorrelation L2 norm regularization pruning technology to maintain the integrity of cross-layer feature fusion path.

Benefits of technology

The lightweight and high-performance infrared small target detection model is achieved, the performance degradation of the model after pruning is avoided, and the stable extraction of small target features and the integrity of the feature fusion path are ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120806025A_ABST
    Figure CN120806025A_ABST
Patent Text Reader

Abstract

The invention discloses a structure grouping de-correlation pruning method and an infrared small target lightweight detection method, which are applied to the field of model compression, a shielding layer is determined from model layers, and a to-be-pruned layer set is constructed based on the model layers except the shielding layer; dividing neurons in the to-be-pruned layer set into a plurality of neuron substructures based on a neuron coupling relationship; dividing the neuron substructures with similar structures into a same neuron substructure set; and determining a to-be-pruned neuron substructure in each neuron substructure set, pruning the to-be-pruned neuron substructure to obtain a pruned infrared small target detection model, and performing infrared small target detection through the pruned model. According to the method, the key structure of the shielding layer protection model is set, based on the coupling dependency relationship of neurons, substructures with strong relevance are identified and formed, groups are formed according to structural similarity, the integrity of a cross-layer feature fusion path is effectively maintained, and the performance of the model is prevented from being reduced after pruning.
Need to check novelty before this filing date? Find Prior Art

Citation Information

Patent Citations

  • Infrared small target detection method for sparse perception global channel pruning

    CN120014408A

  • Large language model pruning method and device, storage medium and program product

    CN120317305A

  • Unmanned aerial vehicle target detection model lightweight method based on pruning algorithm

    CN120633743A