Week 5 — Warmth of Image

Title: Weather Condition Prediction from Image

Team Members: Berk GÜLAY, Samet KALKAN, Mert SÜRÜCÜOĞLU

E-mails Respectively: berkgulay.cs@gmail.com , abdulsametkalkan@gmail.com , mertsurucuogluu@gmail.com

We have approximately 6500 datas. Each data consists of matrix of image with RGB values and we splitted them into training and validation data with %25 ratio.

I built so many CNN architecture and tried them to obtain the best results. I trained images with different sizes like 20×20, 50×50 and 100×100. Besides general accuracy for validation data, I also printed accuracy of each class separately. To do that I split validation data into 5 parts which represents the classes. There are some trials with architecture of network below:

CNN with 6 Layers
Loss Function-1
CNN with 7 layers
Loss Function-2

The last one is the best among the 20 trials.

CNN with 6 layers

Each train data is trained in about 10 minutes. But the last one took 30 minutes. Because its size is 100x100x3. Since there are so many variation, I couldn’t try all. I think this good result for us.

While I use CNN method, my groupmates have tried to extract features from image to use SVM and DT. Most difficult problem for us is to extract feature. We still try to extract feature, because we encounter many errors while extracting feature. But we have obtained some results so far.

Results of Decision Tree:

Trial-1
Trial-2
Trial-3

And result for Random Forest

So, since there are so many variation, we need more time to try other variations.

I hope next week we finish the project and obtain good results for all machine learning methods.


Week 5 — Warmth of Image was originally published in WarmthOfImage on Medium, where people are continuing the conversation by highlighting and responding to this story.

Source: Deep Learning on Medium