I had fun watching this video! Great job. Also very nice to see you making typos and then fixing it. Many people on RU-vid cut those parts and it gives some sort of weird feeling like the person is pretending like they never make mistakes. It's impossible!
Thanks for showing how to use all three methods. For the comparison part I think you could go a lot more in depth on the results. For most applications the results seem largely insufficient to me, but for some applications it is already fine.
Hey Rob, thank you for the Video! this helped me a lot.!!! @to my Indian Developers, I have tried these libraries for Indian languages. And `pytesseract` seems to be a winner fro reading Marathi and Hindi language.
Hi there! I wanted to thank you for the incredible video on text detection in images using `pytesseract`, `easyocr`, and `keras_ocr`. It was exactly what I needed to resolve all my doubts about which tool to use in my projects. Your clear explanation and comparative analysis helped me understand the advantages and disadvantages of each, allowing me to make an informed decision. Thanks again for sharing your knowledge and for the quality of your content. I would love to see more videos like this, where you explain and compare different tools and technologies. Keep up the excellent work! Best regards and well done! Philippe
Hey Rob, thanks a ton for the awesome job you do, I have learnt a lot of new cool stuff (I was only aware of pytessaract, thought it is the best one and tried it on pictures), now I will definitely give a try to the other libraries as well. As I already mentioned, overall you are doing an extremely great job, I just have 1 idea and 1 suggestion : ) 1) idea: as you mentioned in the video, we have the annotations (like the ground truth) already provided in the dataset, right? Wouldn't it be a good idea to check the results vs the ground truth? I.e. smth like lib2 has 3 out of 10 annotations correctly ( recall 0,3) out of 6 predicted texts (0,5 precision) smth like accuracy etc. I understand, this is not perfect as it could only partially extract the text (like miss 1-2 letters) or recognize letter 'G' instead of 'C'. I mean the question is: what's a fair way to have some numerical result based on the dataset? Say library 1 has accuracy X precision Y and recall Z and based on this values and the confusion matrix to be able to make a selection of a library given a particular dataset? 2) Suggestion: I am mentioning this 3rd time already, but your videos are just marvellous, I mean extremely informative, to the point, no second wasted at all. Just perfect. So, what I would really love to see is (as you asked about suggestion for some next video) is almost the same kind of video for some audio task. There is this new BLOOM model out today, I haven't checked it yet, but maybe you could pick some models (like based on wav2vec2) that recognize voice (asr) and compare them using transformers and huggingface for example? Would really be nice to see some comparison of different models in some audio related task, where the whole pipeline (like loading audio, extracting the numerical features, feeding to some pretrained model and finally prediction) would be implemented and validated based on say word error rate. Thanks in advance ; )
Great comment. Wow! Lots to respond to. For #1 that’s a great idea and I had considering doing a more formal evaluation metric to compare the models but ended up deciding it would be too much for this video. I think most of these models are released with their metrics on similar datasets. For #2 I’ll have to check out the model you are referring to I’ve never heard of it before but will read up. I was considering making a video on the new language translation model that meta released last week. Thanks for watching and I appreciate the feedback!
Great video but man I would have loved to see another half hour where you compare the results to the original annotations, give a score to each technique, and then try to make improvements to get the score to increase
Hi Rob, a very informative video thank you! After we extract the data what is the best solution to save it into a table without missing the context of future images or PDFs? Per example if I save Customer's First Name from first file, it should be easier to detect it in the next file and save it under the same column
Hey Rob, Amazing content, thanks for making this. Can you suggest the best method to detect words from a printed text? also like to isolate a single word out of a paragraph. Keep up the good work!
Hi Medallion, do you know how to improve the accuracy of the easyocr ? I have a image , like a dataframe image, it can only read 95% of the data in the image. Thanks in advance
Great question. I’m not sure how you could improve it without training on an additional set of labeled data. You could attempt to use different models and somehow merge the results.
Nice tutorial! I only try pytesseract in the past with PDF files. For future video recommendation I will like to see something about how you organize your ML experiments. In the competitions do you use some tool in order to try multiple models and hyperparameter tuning?
Yes, I think pytesseract is really good for documents, but not so much for extracting text from images. Great reccomendation for future videos. There are a lot of great resources for organizing ML experiments these days. I used to just append all my experiment results to a CSV but now things like Weights & Biases or Neptune.ai make these things a lot easier. I'll have to think about how to make a video for that....
@@robmulla I agree currently trying to extract text from a scanned file. But tesseract isn't getting it right. Picks some words but not all readable words. About to try the others
Thank you Rob for another great tutorial! I've been following along for one of my OCR project. Qq -- when I tried to pip install keras-ocr in Kaggle. I kept getting error message of "ERROR: Could not find a version that satisfies the requirement keras-ocr (from versions: none)". Do you happen to know why? Thanks!
great video! i was asking myself if as a languages i could set "math". i'm trying to get a prediction of number and text together could i use one of this libraries or for handwritten numbers i have to use a totally different library?
Hello Rob. Can you please tell what you think about which one of these 3 will work best if I have to extract text from images of products of inventory? for example soft drinks pictures, chips, snacks, chocolates, tissue boxes, vanity products, etc.
hi, i'm trying to find a way to process bulk ai generated art, and flag any images that contain even the tiniest watermark or text. I fumbled around with some OCR previously in python (don't remember what i used) but it wasn't very close to what i needed it to do considering the text is often just random scribbles or gibberish. any chance you could point me in the right direction to a possible solution? thanks.
Wow nice work... My question is in a situation where they're many text on the image is it possible for me to get just a word.. Can it work by automatic capture and convert to text
Thanks. I’ll see what I can do. The hardest part is finding good training dataset to use! If it exists there is probably a pretrained model that exists.
I am trying to mess with japanese data which is cc100 datasets U can find manga_ocr that's pretty good but i think if somehow we use that data we can make better in EasyOCR for handwriting
@@robmulla I head that it's faster than easyocr, also it has several features: for text detection, text identification, document identification (like tables, ...), mobile and embedded devices support and looks like it's easy to fine-tune (for a new language). I gonna use it and will share experience
can you please video how to prepare our custom dataset with annotation. and how use these dataset with anotation to train a pretrained pytesseract or easyocr or kerasocr
Dear Sir, I am your Subscriber I want to create a tool that finds text errors in the image. For Example: I forgot to write CONTACT US, BUY NOW, CONTACT NUMBER, SPELLING MISTAKE, etc... in my social media post. that the tool finds error and suggests what are missing or what is incorrect in social media post. 🙏 Please guide me and suggest what course I need to buy or what I need to learn to create this tool Thank you
hey can you please make something that can identify how many pages in a PDF are having images and how many pages are non-Image, Thanks in advance, or at least guide me through the process of doing do
Hello sir, I tried to install keras-ocr through pip in kaggle. But it wont install throwing an error saying, "ERROR: Could not find a version that satisfies the requirement keras-ocr (from versions: none) ERROR: No matching distribution found for keras-ocr. What should i do for this?
Hi, I'm a new sub here! I've tried many times to decode a datamatrix code in order to get the text string represented by that code but my Python script seems not able to decode it... I use opencv, libdmtx and pil but nothing... Probably once the script opens the picture and tries to find the code (the pattern) in the image nothing occurs, and it happens well before the decoding effort that should come right after a pattern has been recognized... Please help 😅 Thanks!
Hey. Thanks for subbing. I’m not sure what you mean by data matrix. Are you able to first convert the file to something like a jpg first? That’s what I’d suggest.
@@robmulla So instead of opening it as a png image you rather suggest to first convert it into a jpg image, and then use the decode command... Working with a png image and then using decode generally gives me an empty output []... Thanks! P.s. For the sake of clarity, my png image is a foto of an electronic device with a datamatrix printed on it
I was trying to do EasyOCR to accelerate some software of mine by using GPU instead of pytesseract, but the results are absolute dog shit on obviously readable text. Pytesseract is getting it word for word yet EasyOCR can barely get even a couple of the letters correct. Pretty disapointing, wish pytesseract was GPU enabled from the start.
Great question. It should be possible. I have a RU-vid video about working with video data. Each frame is essentially just an image that you could apply these techniques to. Check it out!
Please help keras-ocr showing error while installing 😢😢😢 Error: couldn't find a version that satisfies the requirement Error: no matching distribution found for keras-ocr
Verify your kaggle profile with your phone number (1 time key/code). Then copy the notebook again, and in the session options (right side of notebood) you will see that "internet is on". This will also let you use GPU resources.
