Enhancing CNNs With Detail Feature Module for High‐Pixel Image Classification

ABSTRACT

In the realm of advanced photographic technology, high‐pixel images offer visual clarity but also present significant challenges to computer vision due to their complex textures and structural details. These high‐pixel images necessitate the development of sophisticated algorithms for image feature extraction to improve classification accuracy. Although convolutional neural networks (CNNs) exhibit powerful capacities for extracting features of images, they lead to rapidly increasing computations. To enhance the efficiency of CNNs, this study proposes a detail feature module (DFM) that employs a spatial attention mechanism (SAM), high‐dimensional weight assignment (HWA), and a feature fusion module (FFM). The DFM can be generally embedded into CNNs, say, residual networks (ResNet) or lightweight ShuffleNets, augmenting the network's ability to capture and retain detailed features by assigning weights from high‐dimensional features to low‐dimensional ones. We carry out experiments on images with varying pixels to assess the effectiveness of the DFM. The comparative and ablation experiments demonstrate that the integration of DFM enhances the classification performance of ResNets and ShuffleNets, with particularly notable improvements in the lightweight ShuffleNets. The DFM proves its generalizability within CNNs and enhanced capability to extract features, underscoring its potential impact on improving the classification tasks in complex imaging environments.