I have a dataset of thermal images resulting from an experiment. Each experiment has 200 images which are in sequence. The images are 16-bit b&w cropped to the ROI. There are some parameters used for the experiment e.g. power levels and phase. I want to construct a model which takes in these sequences of images and outputs the correspoding parameters (8 in total) for the sequence i.e regression.
So far, I have tried
CNN with a regression output. These two approaches have been without any temporal context. In another approach, I constructed a
ConvLSTM, with a normal convolution head without temporal information followed by an
LSTMCell with a regression layer. The loss landscape is better with this approach, but I am not really sure if it is an optimal solution. I also came across the
TimeDistributed wrapper layer of
Keras which can provide temporal information in Linear and Convolution layers.
Contextually speaking, the temporal information is very crucial in my case. I want to know whether I am in the right direction with using something like a ConvLSTM+Regression or TimeDistributed layers.
Some information about the setup I am using for training:
(experiments_count, time_steps, width, height, 1)normalized to
0-1For my case,
Mean Squared Error/
Mean Absolute Error
1since the experiments count is close to 1000 and each image is 300x300