I’m working on a facial expression-based emotion detection project using images collected from different datasets. The challenge I'm facing is that the images have different resolutions, such as: 48x48 (grayscale) 64x64 96x96 128x94 1024x1024, etc. I want to train a CNN model for emotion classification, but I’m not sure how to preprocess the images so they can be fed into a single model. I tried resizing all images to 48x48, since many emotion datasets (like FER2013) use this size. However, when resizing large images (like 1024x1024) to 48x48, the facial details are lost — sometimes the face isn’t even recognizable anymore. What I Want to Know: 1.How should I standardize image sizes from different datasets for CNN input? 2.Is it okay to resize large images to smaller ones (e.g., 1024x1024 → 48x48), or does it lose too much important detail? 3.Would it help to crop the face before resizing? 4.Is there a best practice to handle mixed-resolution images in CNNs? Additional Info: 1.I'm a beginner to deep learning and CNNs. 2.I'm using TensorFlow/Keras for model development. 3.The input data is a mix of grayscale and RGB images. 4.Any advice or examples for a preprocessing pipeline would be really helpful! What I’m Looking For: 1.A recommended approach (ideally with code examples) to: 2.Preprocess mixed-resolution images for emotion detection. 3.Resize without losing critical facial expression information. 4.Handle face detection/cropping, if needed. Continue reading...
