Article Preview
Top1. Introduction
Deep Learning, in simpler terms, is a way to train a model on some dataset so that it could predict the outcome of other similar queries. When the working of these models imitates to that of the human brain, it is termed as Neural Networks.
These Neural Networks have hidden layers and preceptors functioning like the neurons in our brain. This tends to minimize the loss and increase the efficiency of the output results. Neural Networks have shown much promising results in various problems and have seen many advances in past some years, making it one of the top researched fields of computer science.
Images can be in the standard RGB (Red, Green, and Blue) format or can be in greyscale i.e. black and white. It is quite easy to convert RGB Images into Greyscale, but the vice versa have attracted many researches due to its difficulty level. Greyscale Images has only single value for every picture showing the amount of light at that pixel. Thus, it becomes very difficult to know the feature it represents and the colour it will contain in the RGB format.
The colorization problem has paved way for many researches to find an optimal solution for it. Neural networks have shown good results, but it varies on model to model and the training datasets. This paper shows that it is very important to choose the correct parameters in the models and the dataset so that the loss in the final image can be minimized, thus generating an efficient model for the problem.
This paper also develops an efficient model for captioning an image. The image captioning is a classic example of vector to sequence model, where the input is in the form of an image, and the output provides information about the image.
Targeting the problem of obtaining as precise as possible caption of the image, neural networks, or specifically recurrent neural networks, have provided good results in recent times, besides opening scopes for further development. Based on previous researches and studies, this paper tries to generate an effective model, with more accurate parameters, in order to caption the image.
Further now days working on remote systems have also increased for many organizations. In such scenarios, IoT deployment tools come very useful in deploying these neural networks models to various remote systems. Tools like Docker, Greengrass, etc., can be used according to an organization needs and wants. This paper also finds out a comparative analysis of deploying different neural networks models via deployment tools on various functional and non-functional parameters.
The key contributions of this paper are:
- •
This paper shows that the change in the loss rate of a model highly depends on various parameters used in the building of the model;
- •
This paper proposes an efficient model for colorization of greyscale images;
- •
This paper proposes a proficient model for captioning an image;
- •
This paper connects neural networks models with IoT deployment tools;
- •
This paper shows the deployment of neural networks models to remote systems via IoT deployment tools like Docker and AWS Greengrass;
- •
This paper also brings out a comparative analysis of the deployment of different neural networks models via different tools.
The paper is divided into twelve sections. Section 1 is a brief introduction, followed by the objectives of the paper as Section 2. Section 3, 4, and 5 tells about the convolutional neural networks, auto encoders and rectified linear units, respectively. Section 6 elaborates on the colorization models. Section 7 and 8 describe bidirectional LSTMs and Inception V3, respectively. Captioning models are explained in Section 9, followed by deployments in Section 10. Section 11 gives the comparative analysis of the deployments. Lastly, Section 12 provides the conclusion of the paper.
Top2. Objectives
The paper focuses on firstly the colorization of the images, then captioning of the image and then deploying them to remote systems and doing a comparative analysis of the same.