Meat multi-modal basic large model pre-training method based on self-supervised learning

By employing a self-supervised learning method and pre-training a Transformer model using an unlabeled multimodal meat dataset, the problems of high data annotation costs and insufficient model generalization ability in meat detection are solved, achieving efficient and rapid feature extraction and adaptation to various downstream tasks.

CN121561584APending Publication Date: 2026-02-24郭圣峰
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202511760386.0
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-11-27
Publication Date
2026-02-24

AI Technical Summary

Technical Problem

Existing meat testing technologies rely on labeled data, resulting in high data labeling costs and long cycles. The models lack generalization ability, making it difficult to quickly respond to new meat species and adulteration methods, and they lack the ability to extract general features.

Method used

We employ a self-supervised learning approach, utilizing an unlabeled multimodal meat dataset. We pre-train the data using a Transformer-based encoder-decoder model and a masked autoencoder (MAE) to extract the intrinsic structure and features of the meat data.

Benefits of technology

It enables the learning of robust and generalizable general features from unlabeled data, reducing subsequent development costs and time, and improving model performance in scenarios with few samples.

✦ Generated by Eureka AI based on patent content.
Patent Text Reader

Abstract

The invention discloses a meat multi-modal basic large model pre-training method based on self-supervised learning, and belongs to the technical field of artificial intelligence and food detection. According to the method, label-free meat multi-modal data is collected, and self-supervised pre-training is carried out by using a mask automatic encoder (MAE) architecture, so that a model learns general feature representation of meat data. According to the method, the dependence of supervised learning on data labeling is overcome, the obtained basic large model can serve as a powerful feature extractor, various meat detection tasks can be quickly adapted through fine adjustment of a small number of samples, and the development efficiency and generalization performance of the AI model are remarkably improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and food testing technology, and in particular to a pre-training method for a multimodal basic large model of meat based on self-supervised learning. Background Technology

[0002] Currently, AI-based meat inspection technologies largely rely on supervised learning, requiring large amounts of precisely labeled data (such as species, cut, and freshness) for model training. This approach has significant drawbacks: first, data labeling is costly and time-consuming, and depends on the experience of professionals; second, when faced with new meat species, novel adulteration methods, or different processing methods, the model's generalization ability is insufficient, requiring re-collection and labeling of data, making rapid response difficult. Existing models are mostly designed to solve specific tasks (such as species identification), lacking a deep understanding of the general characteristics of meat data, and failing to form a unified and powerful feature extraction foundation adaptable to various downstream tasks. Summary of the Invention

[0003] Purpose of the invention: To address the problems of existing technologies relying on labeled data, poor model generalization, and long development cycles, this invention provides a multimodal basic large model pre-training method that can autonomously learn the essential characteristics of meat from massive amounts of unlabeled data.

[0004] Technical solution: To achieve the above objectives, the present invention adopts the following technical solution: A pre-training method for a large multimodal basic model of meat based on self-supervised learning includes the following steps: S1: Construct an unlabeled multimodal meat dataset, which includes near-infrared spectral data, high-resolution image data, and electronic nose odor data from a large number of meat samples; S2: Construct an encoder-decoder model based on the Transformer architecture as the basic large model; S3: The basic large model is pre-trained using the self-supervised learning paradigm of Masked Autoencoder (MAE). Specifically, this includes: randomly masking a portion of the information in the input multimodal data, encoding the unmasked portion by the encoder, and then reconstructing the masked original information by the decoder. S4: Using reconstruction loss as the optimization objective in the pre-training stage, the model learns the intrinsic structure and feature representation of meat data, ultimately obtaining a pre-trained multimodal basic model of meat.

[0005] Beneficial Effects: This invention, through self-supervised learning, fully utilizes massive amounts of unlabeled meat data to train a large-scale basic model capable of extracting robust and highly generalizable features. Based on this basic model, only a small number of labeled samples are needed for fine-tuning, allowing for rapid and efficient adaptation to various downstream tasks such as species identification, individual association, freshness assessment, and adulteration detection. This significantly reduces the cost and time required for subsequent AI application development and substantially improves the model's performance in scenarios with limited samples.

Claims

1. A pre-training method for a large multimodal basic model of meat based on self-supervised learning, characterized in that, Includes the following steps: Step 1: Collect and construct an unlabeled multimodal meat dataset, wherein the multimodal data includes at least spectral data, image data, and odor data; Step 2: Construct a neural network model containing an encoder and a decoder as the basic large model; Step 3: Pre-train the basic large model using a masked self-supervised learning strategy. The strategy includes: randomly masking the input multimodal data, extracting features of the visible part using an encoder, and reconstructing the masked part of the data using a decoder. Step 4: Optimize the model parameters by minimizing the difference between the reconstructed data and the original data to complete the pre-training of the basic large model.

2. The method according to claim 1, characterized in that, The spectral data mentioned in step one is the near-infrared spectrum in the 900-1700nm band, the image data is a macro or polarized image with more than 20 million pixels, and the odor data is the electronic nose response value with more than 64 channels.

3. The method according to claim 1, characterized in that, The encoder described in step two adopts a Transformer architecture, and the decoder is a lightweight Transformer or a fully connected network.

4. The method according to claim 1, characterized in that, The masking strategy described in step three uses a uniform random masking rate for the three modalities of spectral, image, and odor data, with a masking ratio of 50%-80%.

5. The method according to claim 1, characterized in that, The pre-trained basic model can be used to adapt meat species identification, individual traceability, freshness assessment or adulteration detection tasks through fine-tuning.