ماشین لرنینگ و پردازش تصویر

---

شرح پروژه

Special Topics in Electronics (Machine Learning)

Course Project

Topic: Image classification

Data Set:

The dataset includes 3504 images from the broadcast news domain. This set has been extracted from the TRECVID 2005 development set. Manually labeled ground truths of two concepts “building” and “explosion_fire” for each image are included. There are 555 images labeled as building and 136 images labeled as explosion_fire in this data set. For each image, in addition to the raw pixel values, two sets of image features, 5x5 grid color moments (225-dimensions) and Gabor texture (48-dimensions), have been provided. You should use both raw image data and the features in this project. You can access the dataset from the course website.

Task:

The objective of this project is for you to gain experience in designing statistical classifiers for solving practical problems by applying the knowledge and tools learned in the course. Here are three main steps, which are explained below:

Step 1: You are asked to develop image concept classifiers, which take as input the image features and predict the concept labels.

- Classifying image data using kNN

- Applying a dimension reduction approach like PCA and repeating kNN classification using partial distance method

- Classifying image using both Bayes and naïve Bayes classifier and compare their results and analyze them (Suppose that a unimodal Gaussian model can be used as the conditional probability function)

- Classifying images using SVM with at least two different kernels and 5 different values for the kernel parameter(s) (explain the kernels you have used and the reason of using it and compare their results).

In evaluating the results of your classifiers, please compute the precision and recall values of your classifier. If your classifier includes any parameters that can be adjusted, please report the precision-recall curve showing the performances at different parameter values (for example k in kNN classifier). Remember to partition the data set into K folds and conduct a cross-validation procedure in measuring the performance. For simplicity, keep K small. A two-fold partition will be fine. You should also report execution time of each classifier and compare the results by some explanations.

Step 2: In this step, please use the feature data individually and apply all of the previous tasks on them. Be careful that there are two different feature data available and you should use both separately.

Step 3: Finally, you should compare the final results produced by these three types of data and explain the results.

You have to use “python” language but you are allowed to use any library in python for implementing this project. Please implement your code in “Jupyter notebook” and show all of your results in it.

Extra bonus:

- You can classify the data using a neural or deep neural network. The network and parameter you choose is optional but do not forget to explain the network and its structure and also all the parameters, normalization, loss function and any other setting you have used. Using TensorFlow, Keras or Pytorch library for implementing this network is also optional. You can use pretrained networks but implementing a network have special bonus.

- You are encouraged to investigate advanced topics as well. For example, you may explore fusion of classifiers using different features, models, or families of classifiers. In addition, you may study the techniques of feature correlation analysis and feature selection and analyze their effects on classification performance.

Requirements:

- A project report of 10-15 pages, including text description, analysis, and all the figures. Remember to include explanations about why you choose the specific approaches, your empirical design decisions, and insights you learn in intermediate analysis steps as well as the final results. Please do not include all of your codes in your report and just place some parts of your code that is needed for explaining what you have done. Please provide sufficient details that are needed by future readers in case they want to repeat your experiments.

مهارت ها و تخصص های مورد نیاز

---

---