Study of the Impact of Standard Image Compression Techniques on Performance of Image Classification with a Convolutional Neural Network

In this study, we have measured the impact of image compression on the classification performance of Convolutional Neural Networks (CNNs). By using a pre-trained CNN to classify compressed images, we have shown that on average, an image can be compressed by a factor 7, 16, 40 for a JPEG, JPEG200 and an HEVC encoder, respectively, while still maintaining a correct classification by the CNN. This study also showed that pretrained AlexNet CNN was making use of JPEG artifacts learned during the training phase to perform classification. To further study the impact of compression on CNN-based classification, a large set of encoding parameters was explored: color-space, resolution, Quantization Parameter (QP). Main conclusions of this study are that color is essential for classification with AlexNet CNN, and that classification is resilient to image downscaling. Finally, we have studied the correlation between classification performance of a CNN and image quality measured with two objective metrics, namely the Peak Signal to Noise Ratio (PSNR) and the Structural SIMilarity (SSIM). We have found that the SSIM metrics was more appropriate to measure the degradation of an image with regards the CNN performance.