This chapter reviews existing technologies in the market that do image classification. It also reviews current image classification technologies, which meet the requirements for the project. This review also analyses methods that are used to distribute the learning between various physical machines, and how to spread the load so minimum reloading of the models is required when running the model.
There are currently some existing software as a service (SaaS) platforms that do provide similar services to the ones this will project will be providing.
%Amazon provides bespoque machine learning services that if were contacted would be able to provide image classification services. Amazon provides general machine learning services \cite{amazon-machine-learning}.
Amazon provides an image classification service called ``Rekognition'' \cite{amazon-rekognition}. This service provides multiple services from face recognition, celebrity recognition, object recognition and others. One of these services is called custom labels \cite{amazon-rekognition-custom-labels} that provides the most similar service, to the one this project is about. The custom labels service allows the users to provide custom datasets and labels and using AutoML the Rekognition service would generate a model that allows the users to classify images according to the generated model.
The models generated using Amazon's Rekognition do not provide ways to update the number of labels that were created, without generating a new project. This will involve retraining a large part of the model, which would involve large downtime between being able to add new classes. Training models also could take 30 minutes to 24 hours, \cite{amazon-rekognition-custom-labels-training}, which could result in up to 24 hours of lag between the need of creating a new label and being able to classify that label. A problem also arises when the uses need to add more than one label at the same time. For example, the user sees the need to create a new label and starts a new model training, but while the model is training a new label is also needed. The user now either stops the training of the new model and retrains a new one, or waits until the one currently running stops and trains a new one. If new classification classes are required with frequency, this might not be the best platform to choose.
Similarly, Google also has ``Cloud Vision API'' \cite{google-vision-api} which provides similar services to Amazon's Rekognition. But Google's Vision API appears to be more targeted at videos than images, as indicated by their price sheet \cite{google-vision-price-sheet}. They have tag and product identifiers, where every image only has one tag or product. The product identifier system seams to work differently than the Amazon's Rekognition and worked based on K neighbouring giving the user similar products on not classification labels \cite{google-vision-product-recognizer-guide}.
This method is more effective at allowing users to add new types of products, but as it does not give defined classes as the output, the system does not give the target functionality that this project is aiming to achieve.
\subsection{Requirements of Image Classification Models}
One of the main objectives of this project are to be able to create models that can give a class given an image for any dataset. Which means that there will be no ``one solution fits all to the problem''. While the most complex way to solve a problem would most likely result in success, it might not be the most efficient way to achieve the results.
This section will analyse possible models that would obtain the best results. The models for this project have to be the most efficient as possible while resulting in the best accuracy as possible.
A classical example is the MNIST Dataset \cite{mnist}. Models for the classification of the MNIST dataset can be both simple or extremely complex and achieve different levels of complexity.
For example, in \cite{mist-high-accuracy} an accuracy $99.91\%$, by combining 3 Convolutional Neural Networks (CNNs), with different kernel sizes and by changing hyperparameters, augmenting the data, and in \cite{lecun-98} an accuracy of $95\%$ was achieved using a 2 layer neural network with 300 hidden nodes. Both these models achieve the accuracy that is required for this project, but \cite{mist-high-accuracy} are more computational intensive to run. When deciding when to choose what models they create, the system should choose to create the model that can achieve the required accuracy while taking the leas amount of effort to train.
% TODO fix the inglish in these sentance
The models for this system to work as indented should be as small as possible while obtaining the required accuracy required to achieve the task of classification of the classes.
As the service might need to handle many requests, it needs to be able to handle as many requests as possible. This would require that the models are easy to run, and smaller models are easier to run; therefore the system requires a balance between size and accuracy.
% TODO talk about storage
\subsection{Method of Image Classification Models}
There are all multiple ways of achieving image classification, the requirements of the system are that the system should return the class that an image that belongs to. Which means that we will be using supervised classification methods, as these are the ones that meet the requirements of the system.
% TODO find some papers to proff this
The system will use supervised models to classify images, using a combination of different types of models, using neural networks, convolution neural networks, deed neural networks and deep convolution neural networks.
These types were decided as they have had a large success in the past in other image classification challenges, for example in the ImageNet challenges \cite{imagenet}, which has ranked different models in classifying a 14 million images. The contest has been running since 2010 to 2017.
The models that participated in the contest tended to use more and more Deep convolution neural networks, out of the various models that were generated there are a few landmark models that were able to achieve high accuracies, including AlexNet \cite{krizhevsky2012imagenet}, ResNet-152 \cite{resnet-152}, EfficientNet \cite{efficientnet}.
% TODO find vgg to cite
These models can be used in two ways in the system, they can be used to generate the models via transfer learning and by using the model structure as a basis to generate a complete new model.
\subsection{Well-known models}
% TODO compare the models
This section will compare the different models that did well in the image net challenge.
AlexNet \cite{krizhevsky2012imagenet} is a deep convolution neural network that participated in the ImageNet ILSVRC-2010 contest, it achieved a top-1 error rate of $37.5\%$, and a top-5 error rate of $37.5\%$. A variant of this model participated in the ImageNet LSVRC-2012 contest and achieved a top-5 error rate of $15.3\%$. The architecture of AlexNet consists of 5 convolution layers that are run separately followed by 3 dense layers, some layers are followed by Max pooling. The training the that was done using multiple GPUs, one GPU would run the part of each layer, and some layers are connected between GPUs. The model during training also contained data argumentation techniques such as label preserving data augmentation and dropout.
While using AlexNet would probably yield desired results, it would complicate the other parts of the service. As a platform as a service, the system needs to manage the number of resources available, and requiring to use 2 GPUs to train a model would limit the number of resources available to the system by 2-fold.
% TODO talk more about this
ResNet \cite{resnet} is a deep convolution neural network that participated in the ImageNet ILSVRC-2015 contest, it achieved a top-1 error rate of $21.43\%$ and a top-5 error rate of $5.71\%$. ResNet was created to solve a problem, the problem of degradation of training accuracy when using deeper models. Close to the release of the ResNet paper, there was evidence that deeper networks result in higher accuracy results, \cite{going-deeper-with-convolutions, very-deep-convolution-networks-for-large-scale-image-recognition}. but the increasing the depth of the network resulted in training accuracy degradation.
% This needs some work in terms of gramar
ResNet works by creating shortcuts between sets of layers, the shortcuts allow residual values from previous layers to be used on the upper layers. The hypothesis being that it is easier to optimize the residual mappings than the linear mappings.
The results proved that the using the residual values improved training of the model, as the results of the challenge prove.
It's important to note that using residual networks tends to give better results, the more layers the model has. While this could have a negative impact on performance, the number of parameters per layer does not grow that steeply in ResNet when comparing it with other architectures as it uses other optimisations such as $1x1$ kernel sizes, which are more space efficient. Even with these optimisations, it can still achieve incredible results. Which might make it a good contender to be used in the service as one of the predefined models to use to try to create the machine learning models.
EfficientNet \cite{efficient-net} is a deep convolution neural network that was able to achieve $84.3\%$ top-1 accuracy while ``$8.4x$ smaller and $6.1x$ faster on inference than the best existing ConvNet''. EfficientNets \footnote{the family of models that use the thecniques that described in \cite{efficient-net}} are models that instead of the of just increasing the depth or the width of the model, we increase all the parameters at the same time by a constant value. By not scaling only depth, EfficientNets can acquire more information about the images, specially the image size is considered.
To test their results, the EfficientNet team created a baseline model which as a building block used the mobile inverted bottleneck MBConv \cite{inverted-bottleneck-mobilenet}. The baseline model was then scaled using the compound method, which resulted in better top-1 and top-5 accuracy.
While EfficientNets are smaller than their non-EfficientNet counterparts, they are more computational intensive, a ResNet-50 scaled using the EfficientNet compound scaling method is $3\%$ more computational intensive than a ResNet-50 scaled using only depth while improving the top-1 accuracy by $0.7\%$.
And as the model will be trained and run multiple times decreasing the computational cost might be a better overall target for sustainability then being able to offer higher accuracies.
Even though scaling using the EfficientNet compound method might not yield the best results using some EfficientNets what were optimised by the team to would be optimal, for example, EfficientNet-B1 is both small and efficient while still obtaining $79.1\%$ top-1 accuracy in ImageNet, and realistically the datasets that this system will process will be smaller and more scope specific than ImageNet.
% The models that I will be creating will be Convolutional Neural Network(CNN) \cite{lecun1989handwritten,fukushima1980neocognitron}.
% The system will be creating two types of models that cannot be expanded and models that can be expanded. For the models that can be expanded, see the section about expandable models.
% The models that cannot be expanded will use a simple convolution blocks, with a similar structure as the AlexNet \cite{krizhevsky2012imagenet} ones, as the basis for the model. The size of the model will be controlled by the size of the input image, where bigger images will generate more deep and complex models.
% The models will be created using TensorFlow \cite{tensorflow2015-whitepaper} and Keras \cite{chollet2015keras}. These theologies are chosen since they are both robust and used in industry.
% \subsection{Expandable Models}
% The current most used approach for expanding a CNN model is to retrain the model. This is done by, recreating an entire new model that does the new task, using the older model as a base for the new model \cite{amazon-rekognition}, or using a pretrained model as a base and training the last few layers.
% There are also unsupervised learning methods that do not have a fixed number of classes. While this method would work as an expandable model method, it would not work for the purpose of this project. This project requires that the model has a specific set of labels which does not work with unsupervised learning which has unlabelled data. Some technics that are used for unsupervised learning might be useful in the process of creating expandable models.
While there are various machine learning libraries, the two bigger ones are Tensorflow and PyTorch.
This section will compare the two different libraries.
TensorFlow \cite{tensorflow2015-whitepaper} is an open-source machine learning platform created by Google to develop their production and research systems.
PyTorch \cite{pytorch} is an open-source machine learning library developed by Meta to power their systems.
While both libraries can achieve the same tasks with similar level of accuracy \cite{pytorch-vs-tensorflow-1}, PyTorch is mostly used in research oriented applications rather than applications that might require deployment \cite{pytorch-vs-tensorflow-1,pytorch-vs-tensorflow-2}.
This is generally attributed to the maturity of TensorFlow and TensorFlow's ability to create static graphs, which are optimised for inference.
More important for the project is compatibility with other technologies that the project will use.
In this case, TensorFlow has native support for Go while PyTorch does not.
Which due to Tensorflow's advanced in deployment and compatibility the clear choice for the project.
The technical review of current systems, shows that there are current systems that exist that can perform image classification tasks, but they are not friendly in ways to easily expand currently existing models.
The current methods that exist for image classification seem to have reached a classification accuracy and efficiency that make a project like this feasible.
Model architectures such as ResNet, and EfficientNet have been able to perform image classification on large sets of models and achieve higher than human performances.
Taking these architectures in mind the system should be able to create machine learning models that perform equally well.
As for what technologies to use to build such models TensorFlow seams to be the correct choice as it has better performance when deploying to production, and can more easily integrate with the chosen web technologies.
