WiMi Hologram Cloud Inc., a leading global Hologram Augmented Reality (“AR”) Technology provider, announced that it applied transfer learning to the image classification, and a fusion model for image classification was built to improve the classification performance on small sample datasets by utilizing the feature representation of models trained on large-scale datasets.
Deep transfer learning can apply deep learning models that have been trained on large-scale datasets to new tasks. In image classification, deep transfer learning can accelerate the model training process and improve classification performance by transferring some or all of the network parameters of an already trained model to a new model. The image features are extracted by a pre-trained deep neural network model, classified using a classifier model, the pre-trained deep neural network model and the classifier model are connected, and finally, the whole model is optimized by an end-to-end training approach and by a back-propagation algorithm. This approach can effectively utilize the existing features to improve the accuracy and efficiency of image classification.
A fusion model design is used in WiMi’s deep transfer learning-based image classification fusion model, which combines several pre-trained deep learning models and integrates them by transfer learning to improve the accuracy of image classification. The model architecture consists of the following key components:
Basic model selection: In the design of the fusion model, some basic deep learning models need to be selected as candidate models first. These models are pre-trained models on large-scale image datasets, and they have good performance and a wide range of applications for image classification tasks.
Feature extraction: In order to be able to fuse the different base models, we need to add a feature extraction to each model. The role of this feature extraction is to convert the input image into a high-dimensional feature vector so that subsequent classifiers can classify it. In this feature extraction, we use a convolutional neural network (CNN) for feature extraction.
Fusion: After the feature extraction, multiple feature vectors extracted from the basic model will be obtained. To fuse them, another fusion is designed, the purpose of the fusion is to fuse multiple feature vectors into a more expressive feature vector to improve the classification.
Classifier: Next, a fused feature vector will be obtained. For final classification, a classifier will need to be added, through which the fused feature vector will be mapped to different categories, thus realizing the classification of the image.
Also Read: Landing AI Launches Docker Deployment for LandingLens Models
Fusing the advantages of multiple basic models can improve the accuracy of image classification. At the same time, the fusion model for image classification based on deep transfer learning also has some flexibility, and different base models and fusion methods can be selected according to the actual situation to adapt to different image classification tasks.
Image recognition is an important application of deep learning in the field of computer vision, and the image classification and fusion model based on deep transfer learning researched by WiMi will also be widely used in more industry fields. For example, in the field of intelligent security, the model can be used to perform real-time face recognition on images captured by surveillance cameras, thus realizing automatic alarms for strangers. Autonomous driving is also another important application, where the image classification fusion model based on deep transfer learning can be used to recognize and classify objects such as traffic signs, vehicles and pedestrians on the road. This is crucial for self-driving vehicles to help them determine changes in the surrounding environment and make decisions accordingly. For example, when the vehicle recognizes a pedestrian crossing the road in front of it, it can take timely braking measures to ensure the safety of the pedestrian. In addition, the model can also be used for the vehicle’s automatic parking system, which realizes the automatic parking of the vehicle by recognizing parking spaces and obstacles. In addition, social media analysis is also an application for analyzing and classifying images on social media using the image classification fusion model of deep transfer learning. By classifying images on social media, it realizes the understanding of users’ interests and preferences. For example, by analyzing photos posted by users on social media, relevant products or activities can be recommended to provide personalized recommendation services. In addition, social media analytics can be used for sentiment analysis to understand the emotional state of users by recognizing expressions and emotions in images, thus providing better services and marketing strategies for enterprises.
In addition to the above several application scenarios, the image classification fusion model based on deep transfer learning can also be applied to many other fields, such as smart home, smart manufacturing, smart assistant, etc. By recognizing and classifying images, intelligent perception and understanding of the environment and objects can be achieved, bringing more convenience and efficiency to people’s lives and work.
SOURCE: PRNewswire
