Filtering-out poor-quality images for data preparation

Roopdeep Kaur, Gour Karmakar, Muhammad Imran

Filtering noise is a fundamental part of data preparation that enhances image quality for applications such as object segmentation, detection, and recognition. Various noise reduction techniques are proposed in the literature, including the use of median, Gaussian, and bilateral filters. Convolutional neural networks (CNNs) have gained popularity in image denoising owing to their ability to extract complex patterns and features from data. CNNs are highly adaptable, making them effective tools for various image-denoising tasks. One drawback of CNN-based techniques is that they require an appropriate training dataset and all images to be resized. Another notable drawback of all these filtering techniques is that they work for certain types of environmental and camera noises. To bridge this research gap, in this paper, for the first time, instead of denoising, we propose an approach that filters out poor-quality images for various environmental and camera impacts. In our approach, quality is assessed using an image quality assessment metric and an optimum threshold is used to filter out poor-quality images. We also ensure that a sufficient number of images remain to develop the deep learning (DL) model. The results produced using real and simulated traffic and object recognition data demonstrate the performance supremacy of the proposed approach compared with the state-of-the-art approaches. The average recognition accuracy for our proposed approach is 93.8% for the traffic sign recognition dataset and 84.9% for the object recognition dataset. This indicates our model's potential for real-life applications such as autonomous vehicles.