Scikit-Learn is the most popular Python ML framework, it's a swiss army knife that does a little bit of everything well. Not particularly suited to computer vision, however.
Imbalanced-Learn is a good framework for resampling in imbalanced classification problems. Can be useful if one of your classes is rarer than the other.
Albumentations will help with training set augmentation - this is an important operation on your training set, because it effectively gives you more training data for "free"
OpenCV is a good library for classical computer vision. Note that the author has a... "creative"... opinion on channel order for images, so you'll be tearing your hair out converting from RGB (what most people using this library will be familiar with) to BGR (their opinion, based on an old convention from 24 years ago) and you will shoot yourself in the foot multiple times as a result. Repeatedly, and you won't know why. This deviation from the norm is such a time vampire (Seriously guys, it's not 2000 anymore, why is this still in BGR?!), but the library is also very powerful with a wide range of features, so it's worth the trouble.
Now, ideally you don't reach for a top-shelf tool like a neural network. They're spicy, they're fun, and cool as hell, but if you can get away with a classical approach, you want to try to do that. If it's something like counting cells or objects in an industrial camera, that's the kind of thing that works well here. Stuff where you can come up with a good series of filters, then plug into a classifier or count some pixels. Fiji ImageJ is a good GUI if you want to play with filters without a bunch of coding.
If it's something like "identify the breed of dog generally in social media images with dogs in them" then you'll need a neural network. If you need a neural network, PyTorch is the standard now. More object oriented API, and it's not a Google product (unlike TensorFlow), which is good - Facebook has a good track record of not killing their products.
Finally, one last tip - if you go deep learning, transform your input images from RGB into HSV before feeding into the network. This tends to help the neural network a little anecdotally
xCodehunter
For a project focusing on computer vision,I feel PyTorch is often recommended due to its ease of use, flexibility, and the strong support. it has within the research community. It’s particularly advantageous if you need to experiment with cutting-edge techniques or require dynamic model architectures. PyTorch also offers many pre-trained models and libraries specifically for computer vision tasks, such as torchvision and support GPU and CPU architectures.
horsellama
PyTorch is the mainstream one. Lots of tutorials, docs, libraries, etc
Imbalanced-Learn is a good framework for resampling in imbalanced classification problems. Can be useful if one of your classes is rarer than the other.
Albumentations will help with training set augmentation - this is an important operation on your training set, because it effectively gives you more training data for "free"
OpenCV is a good library for classical computer vision. Note that the author has a... "creative"... opinion on channel order for images, so you'll be tearing your hair out converting from RGB (what most people using this library will be familiar with) to BGR (their opinion, based on an old convention from 24 years ago) and you will shoot yourself in the foot multiple times as a result. Repeatedly, and you won't know why. This deviation from the norm is such a time vampire (Seriously guys, it's not 2000 anymore, why is this still in BGR?!), but the library is also very powerful with a wide range of features, so it's worth the trouble.
Now, ideally you don't reach for a top-shelf tool like a neural network. They're spicy, they're fun, and cool as hell, but if you can get away with a classical approach, you want to try to do that. If it's something like counting cells or objects in an industrial camera, that's the kind of thing that works well here. Stuff where you can come up with a good series of filters, then plug into a classifier or count some pixels. Fiji ImageJ is a good GUI if you want to play with filters without a bunch of coding.
If it's something like "identify the breed of dog generally in social media images with dogs in them" then you'll need a neural network. If you need a neural network, PyTorch is the standard now. More object oriented API, and it's not a Google product (unlike TensorFlow), which is good - Facebook has a good track record of not killing their products.
Finally, one last tip - if you go deep learning, transform your input images from RGB into HSV before feeding into the network. This tends to help the neural network a little anecdotally