Light-Weight Document Image Cleanup Using Perceptual Loss

Smartphones have enabled effortless capturing and sharing of documents in digital form. The documents, however, often undergo various types degradation due to aging, stains, or shortcoming environment such as shadow, non-uniform lighting, etc., which reduces the comprehensibility document images. In this work, we consider problem image cleanup on embedded applications smartphone apps, usually memory, energy, latency limitations device and/or for best human user experience. We propose a light-weight encoder decoder based convolutional neural network architecture removing noisy elements from To compensate generalization performance with low capacity, incorporate perceptual loss knowledge transfer pre-trained deep CNN our function. terms number parameters product-sum operations, models are 65-1030 3-27 times, respectively, smaller than existing state-of-the-art enhancement models. Overall, proposed offer favorable resource versus accuracy trade-off empirically illustrate efficacy approach several real-world benchmark datasets.