Method for improving robustness of deep learning picture recognition algorithm
A technology of deep learning and image recognition, applied in neural learning methods, character and pattern recognition, computing, etc., can solve problems such as errors, manual labeling, and neural network errors, and achieve improved training effects, high recognition success rates, and computational efficiency high effect
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0057] Such as figure 1 As shown, a method to improve the robustness of deep learning image recognition is as follows.
[0058] The system involved in the method includes a training sample set, a training label set, a deep learning neural network, a verification sample set, a verification label set, a test sample set, and a test label set.
[0059] The training sample set is a collection of samples used to train the deep learning neural network. In the process of training the deep learning neural network, in order to improve the training effect, it is necessary to iterate the training sample set multiple times to obtain better training results.
[0060] Each label in the training label set corresponds to a sample in the training sample set, and is used to mark the category to which a certain sample in the training sample set belongs. In the applicable environment of the present invention, the labels in the training label set may be mislabeled.
[0061] A deep learning neural...
Embodiment 2
[0073] Such as figure 2 As shown, a method to improve the robustness of deep learning image recognition is as follows:
[0074] Step 1: Randomly initialize the neural network for deep learning.
[0075] Step 2: Take the training sample set and the training label set as input to the deep learning neural network, and perform forward propagation to output the probability of each category that the deep learning neural network thinks it belongs to. The value of the probability is less than 1:1.2 and the number of iterations of the current sample training set is less than the predetermined value, then put the sample back into the training sample set, and randomly select a sample from the training sample set to repeat step 2. If the category with the highest probability and the probability times If the probability ratio of the large category exceeds 1:1.5, the training sample label is changed to the category with the highest probability.
[0076] Step 3: Backpropagate the new samp...
Embodiment 3
[0080] Such as figure 1 As shown, a method to improve the robustness of deep learning image recognition is as follows:
[0081] Step 1: Randomly initialize the neural network for deep learning.
[0082] Step 2: Take the training sample set and the training label set as input to the deep learning neural network, and perform forward propagation to output the probability that the deep learning neural network believes it belongs to each category. If the probability ratio of the category with the highest probability to the category with the second highest probability More than 1:1.5, the ratio decreases with the increase of the number of iterations, but not lower than 1:1.2, the training sample label is changed to the category with the highest probability.
[0083]Step 3: Backpropagate the new sample and its label to tune the deep learning neural network.
[0084] Step 4: After training a certain number of samples, use the test sample set and test label set to test the training e...
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More