A Multibranch Network With Multilayer Feature Fusion for No-Reference Image Quality Assessment
Wenqing Zhao, Mengwei Li, Lijiao Xu, Yue Sun, Zhenbing Zhao, Yongjie Zhai
Abstract
With the widespread application of digital images in various domains, the accurate measurement of image quality has become particularly crucial. This paper introduces a novel multi-branch multi-layer feature fusion network (MFFNet) to address the inadequate expression of multi-scale and semantic features and local visual feature consideration in existing no-reference image quality assessment algorithms. MFFNet comprises a primary and a sub-branch. Through convolutional neural network feature extraction, the main branch uses a multi-scale feature enhancement (MSFE) module to capture fine-grained features at each layer, thus significantly enhancing its capability to represent local features. It subsequently merges these distinct-scale features through the multi-layer feature fusion (MLFF) module to improve MFFNet performance. Recognizing human attention to the local image area during image quality evaluation, the sub-branch acquires local visual information using a classical superpixel segmentation model. Finally, the two branches are fused using an element-by-element multiplication operation. Comparative experiments are conducted using four representative datasets—CSIQ, TID2013, LIVEC, and CID2013—demonstrating that the MFFNet method outperforms most advanced techniques, thereby establishing the method’s effectiveness.