An Image Pre-Transformation to Suppress Recognition Accuracy Degradation in Compressed Images and Its Analysis

In deep neural network image recognition, it is desirable to use an original image as input to obtain high recognition accuracy. However, when an image is lossy compressed, coding artifacts lead less recognition accuracy. In order to maintain recognition accuracy even for the lossy compressed image, previous works have proposed quantization control methods. However, these methods are not compatible with some encoders because they have large dependence on encoding methods. Therefore, unlike the previous works, we propose an image pre-transformation that maintains the accuracy even for the compressed images with various encoders. A deep encoder-decoder network model is used as the pre-transformation model. This model is learnt with a new loss function that combines recognition loss and the loss that increases the spatial correlation. We evaluate our method with JPEG, JPEG2000, H.265/HEVC and VVC coding standard on ImageNet~2012 classification task. Compared with original images, the bitrates for the transformed images using our method were reduced on all coding standards while maintaining equivalent recognition accuracy. In this paper, we further analyze the signal of transformed images. We found that the transformed images maintain important signals for recognition even after lossy compressed, and reduce intra prediction residuals than the original images.