A spatial-spectral feature fusion attention residual network for automated bacterial classification via atomic force...

Bacterial classification commonly relies on the specialized knowledge and the laborious procedure. To address the issue, a spatial-spectral feature fusion attention residual network (SFFA-Net) is designed to achieve automated bacterial classification in this paper. Specifically, a bacterial image dataset generated by the atomic force microscopy is first established for high-precision network training. Subsequently, the SFFA-Net is built via designing a residual network to perform feature extraction and encode the discriminative representations of the specimens. On this basis, the twin-branch feature extraction module is designed to extract two-dimensional and three-dimensional features, respectively. Subsequently, the improved squeeze and excitation attention blocks are introduced to focus on the key areas of the feature maps to improve the classification accuracy. Besides, an auxiliary classifier is designed to further enhance the model performance by bringing in additional supervision signals. The superiority of the proposed method is verified through comparative experiments.