Challenges involving JPEG format: Connecting concepts from class with the format. Image compression uses a variety of techniques and algorithms, which are recognizably similar to many concepts we discussed in class, like the FFT. Even so, in something like JPEG, the process uses DCTs, which are incredibly similar to FFTs, with the exception of only considering the values, not the phase of the signals. Similarly, the idea of a 2D FFT was only recently touched on, and we have not had much experience translating 2D signals into a new basis. Matlab is also somewhat fickle when dealing with DCTs.
Run length encoding and Huffman coding were also somewhat difficult to classify under the subjects we learned in DSP. They were both intaking signals and manipulating them to be transmitted, but were not exactly a signal and linear time invariant, function of frequency dependent (LCCDE) system like many of the systems learned in class.
Future plans: Finish analyzing the primary methods of Image compression through the encoder, JPEG, PNG, and possibly other methods, and make the connection between these algorithms and the concepts learned in digital signal processing.
Training neural networks requires a lot of time and computational resources. Although we have been able to train a train an autoencoder, it would have been better if we could train it on more samples and change the architecture to work on larger images. Maybe we could find a better trained model online.