Image recognition method and device based on lightweight deep neural network
A deep neural network and image recognition technology, which is applied in the field of image recognition and devices based on lightweight deep neural networks, can solve problems such as high requirements and image classification, and achieve the effect of improving the speed of image recognition
- Summary
- Abstract
- Description
- Claims
- Application Information
AI Technical Summary
Problems solved by technology
Method used
Image
Examples
Embodiment 1
[0054] This embodiment provides an image recognition method based on a lightweight deep neural network, such as figure 1 shown, including the following steps:
[0055] Step 1. Construct a deep neural network model for image recognition; train the constructed model based on the training set to obtain a trained model;
[0056] Step 2. Lightweight network model: Initialize the pruning parameters (parameters of reinforcement learning), including the pruning step size of each feature layer (the pruning step size of each feature layer can be different), and cyclically update the model parameters and pruning parameters; The process of each cycle is as follows:
[0057] First, for the current feature layer in the current model M, use any method to evaluate the importance of channels to sort the importance of each channel, and prune the α channels with lower importance to obtain the model after this round of pruning M'; where the feature layer refers to the module consisting of a con...
Embodiment 2
[0064] In this embodiment, on the basis of Embodiment 1, in the step 2, for any feature layer in the current model M, the importance degree of each channel is calculated based on the scaling coefficient corresponding to each channel in the batch normalization layer. Sorting; the channel with the larger corresponding scaling factor is more important. The channel importance is sorted based on the scaling factor, which greatly improves the sorting speed compared to the sorting based on the feature map.
Embodiment 3
[0066] This embodiment is based on the embodiment 1, in the step 2, for the less important α i Pruning the channels refers to setting the weights associated with the α channels in the convolutional layer, the fully connected layer, and the normalization layer to 0; in addition to resetting the weights to 0, model reconstruction can also be used, that is, Remove the α in the original model i A method of channel-associated structure for model pruning.
PUM
Login to View More Abstract
Description
Claims
Application Information
Login to View More 


