cnn deep learning for low resolution thermal image using transfer learning

Question

I need to train model for object detection for very low resolution thermal camera without data set. The camera resolution is 32x24 pixels, (the camera is MLX90640) Can I use pre-train weigh? or I need to downsample well known data set, and train it from scratch?

What do you mean by "without data set"? The output of a thermal camera is totally different from what pre-trained models are trained on so it is never going to work. You need at least a small annotated dataset of your images to do finetuning. — marco romelli, Dec 10 '19 at 10:22

score 0 · Answer 1 · answered Dec 10 '19 at 09:58

Mobilenet works with a min resolution of 128 * 128 pixels, maybe you can add padding bits or resize(expand) image and perform object detection but that may not get good results.

As the resolution is almost the size of mnist dataset you can try using all algorithms used on it.

cnn deep learning for low resolution thermal image using transfer learning

1 Answers1