Our team has done research after taking some insight from open source work in Malaria detection, different Blood cells detection etc.
We used Object Segmentation CNN deeplearning technique to train the model. We used RCNN for it. Annotation was outsourced as we needed pixel based boundary for each cells in Images present to train our model.
We wanted to explore Image enhancement as part of our previous projects which had poor quality images. Increasing the resolution and making the images more clear will increase the image classification accuracy and accuracy of other image based deeplearning algorithms.
We trained a GAN based Deeplearning model to increase image resolution and also increase the clarity of images by removing blurriness, speckles and denoising them. Trained data was created by adding noice to clear images using small OpenCV based image processing, which involved adding speckle, noise layer by playing with Alpha channels.
Deep learning Object segmentation algorithm Mask RCNN was used to segment different objects based on category. Category masking requirement was pixel wise rather than bounding box, so Mask RCNN worked well in this case.
We also used the same pipeline for detecting required IDs in group of scanned documents , detect accurate boundary, classify each ID documents. To increase accuracy we used Augmentation techniques so as to train model and make it more robust.
We used Face Detection that identifies human faces in digital images. It has various applications like attendance tracking, person recognition etc.
One shot learning was used in it. Facenet model was used that uses triplet loss to determine the similarity bw image embeddings. Embeddings are calculated based on Deep learning method. Initial Person registration in the system happened from the images of ID photos.Then the facenet model used to detect the face embedding using eucledian distance to classify into any person. We then used the person ID to get other details, tracking attendance etc. This system can easily be extended to track childrens, analyse their behaviour etc.
We performed OCR on KYC documents : PAN , Aadhaar, Passport etc to extract user information. To create our own OCR pipeline we used CNN based text localisation with Tesseract for Text classification. Tesseract uses BiLSTM based model to classify texts.
A pre processing pipeline was created to detect ID cards classification and different rotated order so as to align it perfectly for OCR to be performed accurately.
In this project we did some Tag extraction based on the Product Description using Word2Vec based model. Hashtags were generated based on entities extracted from product description.
We can apply same methodology in generating Automatice tags from any content, be it social media post, description of any item, paragraph etc. The model would determine key entities and make hashtag out of it.This can be used for indexing such contents and retrieval algorithms can be built on top of it.