This repository holds my work on computer vision technology.

Projects:

1. Image classification:

• For this project i used Intel Image Classification dataset, which contains image data divided in 3 sub-folder seg_pred,seg_test and seg_train.

• This is a 6 class classification project where i have 6 different classes of images that is mountain, street,buildings, sea, forest, glacier.

• For further explanation please go to project folder and for code implementation visit my kaggle notebook.

---

2. Style Transfer with Deep Neural Networks.

• This project is based on Leon A. Gatys’ paper, A Neural Algorithm of Artistic Style.

• Neural style transfer is an optimization technique used to take three images, a content image, a style reference image (such as an artwork by a famous painter), and the input image you want to style — and blend them together such that the input image is transformed to look like the content image, but “painted” in the style of the style image.

• For detailed explanation please visit my kaggle notebook.

• UI ready project:- click

---

3. Faster-rcnn for multiclass object detection.

• R-cnn and it counterparts were the state-of-the-art object detection algorithms for the longtime until YOLO comes into picture.

• In R-cnn paper Ross girshick first introduce the Region Proposal technique. In which he proposed the idea of using selective search algorithm to extract just 2000 regions from the image instead of classify a huge number of region.

• Faster-rcnn is the updated version of R-cnn where some of the drawbacks are solved. In which the apparach is similar to r-cnn but instead of feeding 2000 regions we feed input image to CNN to generate feature maps.And from these feature maps we identify region propasals, and by ROI pooling layer we eshape them into a fixed size so that it can be fed into a fully connected layer. From the RoI feature vector, we use a softmax layer to predict the class of the proposed region and also the offset values for the bounding box.

• For this project i used PASCAL VOC 2012 dataset, where we have annotations(metadata of images in the xml file) and sets of images. We have 20 different objects of labels and per image can contain 0 to 20 different objects.

---

4. Image Similarity

This project is based on research paper name Classification is a Strong Baseline for Deep Metric Learning.This paper shows the effectiveness of classification based approaches on image retrieval datasets.

Points to note:

• For effective training we make sure that every mini batch contains same amount of classes for per batch, in case of CUB-200 ,we will use 75 image per mini batch with 25 images for per classes(3 classes per batch).

• We will use custom NormSoftmax loss function as author of the paper suggests.

• We will use Layer normalization just before softmax activation layer.