I study German and decided to write app to simplify some tasks:
1) generate anki cards from text
2) extract highlighted words from paper and generate anki cards
3) collection of texts/dialogs on random topics
4) and so on... )
not sure if someone needs it, but very helpful for me )
Technologies: Python backend stack, some computer vision, PDF processing (split, merge, ocr, text extraction, fixing all weird issues, mupdf, ghostscript - worked on pdf related projects), AWS
Doesn't work very well.
First of all there are several separate problems: 1) detect if OCR is required 2) image optimization 3) preprocessing of broken pdf files. And all of them are not easy:
1) page could contain selectable text, but text can't be copied because embedded font doesn't contain glyph->symbol code mapping. Mapping table could contain complete garbage. Sometimes page could contain long urls (added by email services) but all text is provided as image. Sometimes text contains normal text and garbage. And many many other cases.
2) some old scanners generate pdf documents built from 2-5 pixel image stripes. Some of them try to do OCR (poorly). Some of them uses huge DPI. Sometimes you get uncompressed doc in which each page could take up to 200mb. So you need to convert pdf page to image. But you have to choose format and compression options. PNG is ok, but you have to choose correct options (for ghostscript). But output image will be huge. JPG is better, but quality could be low. Sometimes multistage optimization is required. Also tools like ghostscript, fitz or imagemagic doesn't handle all possible pdf/image.
3)weird pdfs - endless story. Poor fonts, broken fonts, very specific cases in pdf standard, issues with image extraction, table of content, viruses, embedded files, annotations, margins/paddings/rotations/translations.
Technologies: Deep Learning/Python Backend stack - keras, pytorch, Flask, Docker, Kubernetes and so on. Fields - image processing, face detection, face recognition, object detection/classification, segmentation.
It detects faces, objects, tags, extracts metadata and provides search interface and API. So you can for example find image with "cat and dog near river with some person"
I want to build enterprise level image search :)
Google Images, for example, uses in most cases alt text and surrounding content.
Google Photos is limited to personal collections.
I have finished prototype and now trying to convert it to startup
1. Khumbu is an image search engine based on computer vision and deep learning.
2. Workflow: upload your images via API or web app and run search queries via API/app.
3. Right now it provides object detection, classification, exif extraction, face detection, nsfw scoring.
4. Tech details: images are processed using celery workers, after that all information is moved to ElasticSearch to serve user queries. Most part is deployed in k8s cluster, but some processing is done on remote machines with GPUs.
Soon we want to add:
1. face recognition (public and user specific faces)
Small self-promotions: We are working on image search engine - https://khumbu.im, demo (MVP stage)https://app.khumbu.im/search/5dff72e66483e25b40e0222e . We use DL to understand content of images and provide better search results. Soon we are going to add face recognition, improve object detection and semantic analysis.
I am working on image search engine based on AI (demo: http://demo.khumbu.im/ ). Khumbu detected faces, objects, scenes, metadata on images etc and provides flexible search using this information (see http://demo.khumbu.im/examples/). UI is a little bit ugly(sorry, I am backend person), but everything else is up and running. Currently I work on REST api, k8s deployment and some AI-related things.
I also got some small investments to rent servers.
I am looking for person who could work on UI related things, product improvement and has experience in startups. Europe is preferable.
I am working on image search engine (something like elasticsearch but for images). I apply some deep learning magic (like object detection, classification etc), face recognition, exif extraction and some other. MVP is almost ready and soon I will be able to handle quries like:
content="dog and cat on the table" exif.apperture>5 exif.data>2005 faces.count >4
But I am pretty weak at frontend development and pitching/presentations.
email: [email protected]