2.1 Image Classification¶
2.1.1 What is Image Classification?¶
Image Classification is one of the most popular application of computer vision. The idea is quite simple and there are many techniques to help us achieve it. While I won’t be covering the various techniques which help us achieve Image Classification, I will be giving you the higher level overview of the concept. As the name suggests, given an image we want to classify in which category does it belong to or in more subtle terms, describe a picture in one or two words. The algorithm which help us achieve Image Classification are Convolutional Neural Networks (CNNs). This algorithm when tweaked and modified can lead to various architectures with different performance scores. Over the period of 6-7 years, a lot of research is done in this domain. Some popular architectures using CNNs as their core component include - Residual Networks (ResNets), GoogleNet, VGG-16, etc.
Intuition of a Binary Image Classification Problem
credits: becominghuman¶
2.1.2 Applications of Image Classification Algorithms¶
We come across image classification algorithms in our daily lives and yet we don’t really know that we have been using them since a long time. The biometric face recognition and Iris classification on your mobile phone is nothing but Image Classification under the hood. Figuring out whether a face is the real owner or the Iris is the one registered on device or is it someone new? - such cases are tackled with human level accuracy by our phones. Ever used Google Photos? If you must have observed, Google can segregate your photos based on your experiences and based on people. A simple search of mountains can bring up all the pictures on your phones where there are mountains. Or if you miss your mom, simply typing Mom will show up pictures of you with your mom. All these are applications of Image Classification, and they have been around for quite a while.
2.1.3 Challenges faced by Image Classification Algorithms¶
One important thing to note with respect to Image Classification is that, given an image, it does not localize any particular object in the image. What do we mean by localize? By localization, we mean the location of the object in the image. An object can be present in the left half of the image, the right half, the center, or the top or bottom, etc. It can be anywhere in the image. As long as it is in the image, the classification algorithm works efficiently. The caveat of classification algorithm is that it cannot be used for detecting object locations relative to image or for figuring out multiple instances of the same object in the image. Such problems are addressed by Object Detection algorithms.
Image Classification vs Localization with Classification
credits: technobezz¶
