Fused features for no reference image quality assessment

Computer vision algorithms for image classification, object detection or segmentation tasks require high-quality images for accurate predictions. With quality degraded images, the algorithms may not detect objects properly or can lead to false detections and wrong classifications. To avoid such problems, we developed an image quality estimator to assess the quality of input video feeds for deciding whether to use the frame or need enhancement or discard it altogether. Traditional algorithms fall short of neural networks in terms of accuracy when estimating image quality. Despite remarkable progress in image processing tasks, Neural networks have not been used widely for image quality estimation. The reason might be the lack of large datasets for image quality assessment, which are subjective, expensive, and time-consuming to create. We propose a Fused Feature Image Quality Estimator (FIQE), which uses both traditional handcrafted and convolutional features to estimate the image quality. The performance evaluation results are compared against the state-of-the-art methods and obtained LCC score of 0.956 and SROCC of 0.955 on the LIVE Dataset and 0.922 LCC and 0.904 SROCC on the TID 2013 Dataset for the regressor model. FIQE classifier model achieved 71.06% top-1 accuracy and 97.46% top-2 accuracy on the LIVE dataset.