Evaluation and optimization of image compression for Convolutional Neural Network in segmentation and classification

The cloud computing for computer vision is in expansion but the use of videos and images for these applications will need an important bandwidth of the network. In order to reduce the bandwidth and the memory used by this videos and images, the use of standard codecs is privileged but these codecs were optimized for the Human Visual System and therefore can be re-optimized for computer vision applications. Furthermore, the neural networks are usually trained and tested with full quality images but the use of lossy compressed images would diminish the bandwidth require. It is therefore interesting to evaluate the impact of the use of these images and to optimize the networks for these images. Material and methods: To evaluate these different approaches we tackle two computer vision problems resolved by deep convolutional neural networks: the multi-organ segmentation of the male pelvic region on CT and CBCT images and the classification of the CIFAR-10 dataset. For the multi-organ segmentation, we first assess the 3D and 2D U-net robustness against JPEG2000 compression before optimizing the networks for various compression ratios by retraining them with compressed images. For the CIFAR-10 classification, we train a standard network with uncompressed images before optimizing the JPEG quantization tables to maintain for each rate the best classification accuracy available. To do so, we use two different algorithms: a greedy one related to the Lagrangian technique and a handcrafted genetic algorithm. Results: For the first task, we find that the U-net multi-organ segmentation robustness to JPEG2000 can be improved by selecting the right network and by fine-tuning the network weights with a new training with compressed images. We show that for a compression ratio of 128:1 we can nearly double the segmentation quality (Dice coefficient) between the worst configuration (2D-network) and the best one (3D-netwtork with fine-tuned weights). For the second task, we find that by using the optimized quantization in JPEG the classification accuracy can be improved by up to 3%. Conclusion: The compression and the DL architectures can be optimized in order to compress the image to a higher level while maintaining equivalent DL performances.