Computer vision is a sub-field of Artificial Intelligence that helps machines to understand visual representation by analyzing digital images. This is achieved by training different deep learning algorithms on image datasets(or video) to recognize different patterns in the image pixels. There by giving them ability to see and accurately identify objects and classify them into a certain category. In simpler words computer vision includes a set of algorithms which help machines see the real world and make decisions based on their vision.
Transfer learning is an active research topic in the field of machine learning. As the name suggests transfer learning is the art of conserving the amount of knowledge gained in solving one problem and then applying this saved knowledge to a new problem and doing a better job at it. This is a very important break through in the field machine learning and data sciences as data is the only fuel that powers machine learning algorithm. With the help of transfer learning even if the data is less we can still acquire the best results.This reduces the need for data related to the specific task we are dealing with. This reduces the need for data related to the specific task we are dealing with. For this an example would be suppose a model which was earlier used for recognising bicycles will now recognise bikes.
The image above clearly show how Transfer Learning actually works. We can see that in the second step same model which was used to create solution for task 1 has being transferred to solve task two. This way the model before starting on problem second already has some prior knowledge about it. Also it should be kept in mind that Data2 in this case represents a very small amount of new data that the model hasn't seen yet, kinda like the test set for this.
Different research teams across the world have come up with multiple state of art models for the purpose of performing transfer learning only and the best thing about this is these research groups have made most of there models as open source, so anybody with slight understanding of the subject matter can download these models and you it for there own purposes. Transfer learning can also be applied to NLP problems just as easily it is applied to computer vision problems, but that is beyond the scope of this post so we will leave that aside for now. Some of the most prominent pre-trained model used for transfer learning in computer vision are listed below :
Alexnet
vgg16-19
ResNet
GoogLeNet
SqeezeNet
InceptionV3,
ResNet,
MobileNet,
Xception,
InceptionResNetV2
Densenet
Facenet
Deep Transfer Learning
Deep learning as attained enormous amount of fame in the past decade. With the help of this technique we are now able to deal with complex problems which seemed to be impossible to the previously. The only draw back the previously used deep learning model had was the needed data in large quantities in order to over come the problems at hand. However with the advent of transfer learning , deep learning has just pasted it biggest milestone. There are many deep learning models with top-of-class performance that have been developed and tested across domains such as computer vision and natural language processing.
These models are trained on large data sets once the training is complete the model weights and state are saved. These same models could then be used in collaboration with a new simple model in order to get better results on lesser amounts of data. The pre-trained model in this case becomes more of a feature extractor then a classifier or regressor. Once the features are successfully extracted then a new model with simple architecture is then used. This new model takes input the features extracted by the pre-trained model and makes predictions.
The following image shows the clear criterion between a traditional deep learning techniques and Transfer Learning techniques.
One should notice in the above image that the target variable in case of transfer learning,
contains a label set which is a sub set of the label set the pre trained model was originally trained on. Some of the most commonly used frameworks for transfer learning include :
Pytorch
Keras
Tensorflow
theano
Pillow
numpy
pandas
openCV
jupyter notebook
Few applications of computer vision with transfer learning are given below:
Real world Simulations
gaming
Image Classification
Zero Short Translation
Text classification
Sentiment Analysis
Face Recognition
Object Recongnition
Gesture Recognition
Object tracking
Feel free to contact us on contact@codersarts.com in order to get any kind of assignment help | project guidance on the above mentioned topics.