Image classification using transfer learning and Pytorch

In this article, we’ll be discussing on how image classification using transfer learning and Pytorch library helps developers to train deep neural networks with comparatively low amount of data. As we know, image classification is an integral part fo many businesses like e-commerce, automobile, stock photography, social media, medical image analysis etc. The objective of image classification is, as the term suggests, to classify images according to a common theme.

For example, a stock photography website cold have thousands of stock photos. They need to classify these photos into different categories like art, animals, nature, sports etc and again classify into sub types such as dance form photos, horse photos, puppy photos and so on. There can be hundreds of categories, sub categories etc. Image classification helps us predict what an image represents, and classify an image or a databse of images into different categories. 

Accurate classification of images help the end user to easily find the image they are looking for. Image classification is mostly done by manual process. Either people who uploads the photo are asked to tag the photos or the business would employ humans to classify the images, which is of course tedious and monotonous job.

Social media websites is another area where image classification is extensively done. Today, these companies rely on artificial intelligence to do the image classification.  A good example of this would be Facebook, which can identify a person’s face from an image with up to 98% accuracy, with only a few tagged images and classify it into the user’s album.

Various companies in different sectors, such as retail, automotive, healthcare, and defense, are adopting image recognition technology. According to a recent research, the image classification market is estimated to grow from USD 15.95 Billion in 2016 to USD 38.92 Billion by 2021, at a CAGR of 19.5% between 2016 and 2021. 

Image classification using AI

Using AI models, we can easily automate the process of image classification. Here, an image classification model can be trained to identify various information classes. These classes can be anything, such as types of vehicles, human faces, etc. For example, consider a model that has been trained to identify three classes— bike, truck, car. When we add a new image, the model will provide the probabilities of the image representing each of the types of vehicles it was trained on. Consider that you have provided the model with an image of a bike, the expected output would be something like this-

This output would show that the image has the highest probability of representing a bike. 

Image classification – Use cases

• Automotive industry: Advancements in technology are pushing towards the development of self driving cars. Such vehicles should be able to detect obstacles and warn drivers about proximity to guardrails, walkways, other vehicles etc. Image classification technology can help achieve this along with reading road signs and stop lights.
• Drones: Drones that are embedded with image classification technology can deliver vision-based automatic monitoring, inspection, and control of the assets located in remote areas.
• Military surveillance: Enables the detection of suspicious activities in the border areas and automatic decision-making capabilities
• Manufacturing: Image classification can help in inspecting production lines, evaluating critical points on a regular basis within the premises, monitoring the quality of the final products to reduce the defects, assessing the condition of the workers etc.
• Visual search engines: Search engines can be used for visual searches. Users can simply drop an image into the search bar and the image classification technology will automatically provide the most relevant results

A case study of Image classification

One of our clients, Monger is an e-commerce company based in the USA. The client wanted to reduce the manpower required for product image classification. And for that, they wanted a solution that could help them classify thousands of product images with a high degree of accuracy. We developed an AI-powered image classification tool using machine learning algorithms.

The AI-based Inventory classifier tool compares two different product images and finds the degree of similarity between the two. The tool also compares the names of the different products to calculate their differences. In the end, a final score is assigned by the system. Using this score, the system can classify the product. It can even conduct product name analysis and tag them with appropriate category names. Through image classification, the client was able to sift through thousands of product images and sort them quickly and accurately. 

Read full case study – AI-powered image classification tool

Image classification using transfer learning

Transfer learning technique allows developers to train deep neural networks with a comparatively low amount of data. It involves the reusing of a pre-trained model for a new problem. The majority of AI models that solve complex problems require a lot of data, and getting large amounts of labeled data for supervised models can be extremely challenging, considering the time and effort it takes to label data points.

In such scenarios, transfer learning can be extremely useful as it takes existing knowledge and applies it to solve a new problem. Depending on the type of traditional ML algorithms involved, transfer learning can be categorized into three- 

• Inductive transfer learning: The source and target domains are the same, yet the source and target tasks are different from each other.
• Unsupervised transfer learning: The source and target domains are similar, but the tasks are different.
• Transductive transfer learning: There are similarities between the source and target tasks, but the corresponding domains are different. Here, the source domain has a lot of labeled data, while the target domain has none.  

Image Classification using Transfer Learning and Pytorch

Pytorch is a library developed for Python, specializing in deep learning and natural language processing. It use Graphical Processing Units (GPUs) to implement a deep neural network faster than training a network on a CPU.

Recently, Hackernoon launched a competition for image classifiacation that caught my attention. The challenge was to classify a set of images, consisting 364 images of different Indian Dance forms. I participated in it and reached the top of the leaderboard. I’ve written a detailed article with steps, explaining how to improve image classification with less amount of data using pytorch. The notebook for this project is available in my Github. If you need any help with image classification in your business, drop us a message.