International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 12 Issue: 08 | Aug 2025
p-ISSN: 2395-0072
www.irjet.net
Fusion-based Blind Image quality assessment for wild images Vanashree P S1, Lalitha S2 1BMS college of Engineering, bengaluru 2BMS college of Engineering, bengaluru
2 Associate Professor, Dept. of Electronics Engineering, BMS college of Engineering, Karnataka, India
---------------------------------------------------------------------***----------------------------------------------------------------------
Abstract - An efficient convolutional neural network;
has real-time perceptual quality score prediction and incorporates a backend for computation of optional traditional no-reference metrics like BRISQUE, NIQE, and PIQE [9], [12]. The predictions are compared to Mean Opinion Scores (MOS) of the Challenge DB dataset [2], [6], which,shows high correlation with human judgements at a low computational cost;suitable for deployment on mobile platforms.
(CNN)-based approach to blind image quality assessment (BIQA);for real-world distorted images is proposed, optimized for real-time execution on mobile devices. The framework is learned on the Challenge DB dataset to estimate perceptual quality scores without reference images. Inclusion of global average pooling in the architecture reduces computational overhead with predictive accuracy. The trained model is also converted into TensorFlow Lite (TFLite) and added to an Android app that can perform on-device quality prediction. The system is tested on the Challenge DB dataset based on MOS vs predicted quality score correlation. Experimental results have strong correspondence with human vision, reaching a Spearman's Rank Order Correlation Coefficient;(SROCC) of 0.83, a Pearson Linear Correlation Coefficient;(PLCC) of 0.86, and a Mean Squared,Error (MSE) of 0.021. With model size as low as 0.39 MB, the system is properly adapted to resource-limited environments and real-time applications.
1.1 Overview of Proposed BIQA System The explosive development of mobile imaging technologies has resulted in the growing need for accurate and real-time quality assessment tools [3], [5]. Previous BIQA models, although precise, are computationally expensive and not ideal for real-time use on smartphones. To fill this gap, the system to be proposed uses a quantized and optimized lightweight;CNN model that can be used for TFLite deployment [4], [14].The Android application that has been developed offers an easy-to-use interface that enables users to capture images in real-time or choose images from the device gallery. The application predicts perceptual quality in real-time and, when integrated with a backend, retrieves other quality metrics such as BRISQUE, NIQE, and PIQE to perform comparative analysis [12], [20]. This two-layered system design makes the tool efficient, scalable, and easy to use, making it a good choice for tasks such as photography improvement, visual content analysis, and automated low-quality image filtering.
Keywords: BIQA, CNN, Deep Features, Image Quality, Real-World Images, Quality Metrics, Batch Processing.
1.INTRODUCTION Blind Image Quality Assessment (BIQA);has become an important research area in recent years because of the explosive growth in multimedia content taken and distributed on mobile platforms. Conventional Image Quality Assessment (IQA) methods were based on reference images, which in real-world situations are usually not available [1], [8]. BIQA sidesteps the requirement for reference images by learning to estimate perceived quality directly from distorted images.
1.2 Literature Review Initial BIQA work concerned manual statistical features derived from natural scene statistics (NSS). Sadiq et al. [19] employed stationary wavelet transform-based NSS for quality prediction but were not robust in difficult realworld situations. The advent of deep learning has seen various models enhance the accuracy and generalization of BIQA. Yang et al. [10] improved feature representation with data-driven transforms, whereas Song et al. [9] introduced iterative training for realistic distortions. Graph-based models such as GraphIQA [15], [18] introduced graph-structured feature representations that enhanced distortion-specific sensitivity.
Recent research on deep learning has greatly;enhanced BIQA performance by employing convolutional neural networks (CNNs), transformers, and attention-based designs [5], [11], [13]. These approaches have facilitated more effective modeling of hard-to-characterize distortions like blur, noise, compression artifacts, and lowlight environments. Yet, efficiently integrating these models into resource-limited platforms like smartphones is still a problem. Here, a light CNN-based BIQA model implemented on an android app with TensorFlow Lite (TFLite). The system
© 2025, IRJET
|
Impact Factor value: 8.315
|
ISO 9001:2008 Certified Journal
|
Page 437