General approach to developing an image classification algorithm for Dilbert cartoons

Question

As a self-development exercise, I want to develop a simple classification algorithm that, given a particular cell of a Dilbert cartoon, is able to identify which characters are

Answer 1

This question was asked 5 years back, therefore the answers provided above are outdated given the fact that deep learning has changed the face of computer vision over the past 3-4 years. A deep learning based solution would involve training a Convolutional Neural Network, which would learn features and perform classification in an end-to-end learning framework. However, since multiple cartoons may be present in the same image, the standard softmax cross entropy loss used in image classification may not be appropriate. Hence, independent logistic regression should be used as a loss function. Threshold for each class can be obtained based on accuracy obtained over a held-out validation set. Even for cartoons, it is better to use a pre-trained model initialized using imagenet instead of training from scratch (https://arxiv.org/pdf/1611.05118v1.pdf, although the final task in this paper is different, they still do processing on cartoons). If you have abundant data, pre-training may not be that important. This task can be performed using standard libraries like caffe/torch etc.