3D Image Recognition Training with Pre-Trained Convolutional Blocks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for training 3D image recognition models are inefficient, requiring a long time and resulting in low model training efficiency.
Innovation Solution
Pre-train a second recognition model using cubes extracted from 3D images by rotating and sorting N target cubes, then use the convolutional block of this model as the convolutional block of a first recognition model to improve training efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a 3D recognition model is trained using a large quantity of 3D image samples, then the recognition accuracy is improved, but the training time increases significantly resulting in low training efficiency
Solution Approach 1:
The patent applies preliminary action by pre-training a 3D convolutional neural network model using a large quantity of 3D image samples beforehand. The pre-trained model's convolutional features are then transferred to the target recognition model, eliminating the need to re-train these features from scratch. This preliminary preparation significantly reduces the training time for the final recognition model while maintaining high recognition accuracy, as the model inherits learned features from the pre-training phase.
2Reliability
If the model is trained from scratch using traditional methods, then the model can achieve good recognition performance, but the training process takes a long time
Solution Approach 1:
The patent applies parameter changes by modifying the training approach from traditional full-model training to transfer learning with frozen convolutional layers. The key parameter change is in the training strategy: instead of training all model parameters from scratch, the patent freezes the convolutional parameters from a pre-trained model and only trains the classification parameters. This parameter modification maintains recognition performance while dramatically improving training efficiency.
3Loss of information
If traditional 3D model training methods are used, then comprehensive feature learning is achieved, but the training complexity and time consumption increase
Solution Approach 1:
The patent applies the extraction principle by separating the convolutional feature extraction component from the classification component. The convolutional neural network's feature extraction capabilities are extracted and pre-trained independently using large 3D image datasets. This extracted feature extractor is then transferred to the recognition model, where only the classification layer needs training. This extraction approach maintains comprehensive feature learning while simplifying the overall training process and reducing complexity.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
An image identification method and device, an identification model training method and device, and a storage medium. The image identification method comprises: obtaining a target three-dimensional image to be identified (S202); inputting the target three-dimensional image into a first identification model (S204), wherein the first identification model is used for identifying the target three-dimensional image to obtain an image type of the target three-dimensional image, a convolution block of the first identification model is the same as a convolution block of a second identification model for identifying the target three-dimensional image, and the second identification model is a model obtained by training an original identification model by using a target training sample, wherein the target training sample comprises cubes obtained by rotating and sorting N target cubes obtained from a three-dimensional sample image, and N is a natural number greater than 1; and obtaining the image type of the target three-dimensional image output by the first identification model (S206).