Automatic JPEG Compression Using a Color Visual Model

JPEG compression is extensively used in digital cameras, internet, and image databases. The amount of compression can be adjusted by scaling the quantization table or Qtable. In many cases, an iterative process is used to achieve optimum compression having a smaller file size, but still visually lossless at the intended display and viewing distance. This process is very time consuming for large image databases since human observers are used to judge the image quality. We present an automatic method to achieve the optimum compression using a color visual difference model (CVDM). The CVDM output is a map of the visible differences between reference and distorted images. In order to use the model in automatic compression, a single number JPEG artifact score was derived from the visual difference map to be used as a merit function. A subjective experiment was conducted to find the best merit function for JPEG artifacts. The subjective experiment also derives an acceptance criterion for JPEG artifacts. We found that the 99-percentile provides the best correlation with the subjective results; thus it was used as the JPEG artifact score in the automatic compression. In the compression process, the compressed images were evaluated with the visual model. Based on the predicted artifact score, the selected Q-table was scaled up or down so that the artifact score was close to the acceptance criteria derived in the subjective experiment.