Arnet: research on wood CT image classification algorithm based on multi-scale dilated attention and residual dynamic convolution

Addressing the challenges of low classification accuracy and protracted identification times posed by lightweight convolutional neural networks (CNNs) for wood micrograph classification, this study introduces ARNet, a novel model tailored for wood CT image analysis.ARNet significantly enhances the overall image recognition performance by boosting its dynamic feature extraction capabilities and refining its proficiency in processing salient features.The methodology employs residual dynamic convolution that dynamically aggregates convolutional kernels in response to the input image, optimizing adaptability.This optimized field of view across disparate feature layers facilitates the extraction of critical information such as wood texture, pore distribution, and cellular arrangement, thereby enhancing analytical depth.Additionally, ARNet incorporates multi-scale dilated attention mechanisms to capture nuanced feature maps across multiple scales, thereby broadening the scope of feature analysis.This approach not only achieves a profound understanding and efficient processing of the input data but also accentuates critical features, significantly enhancing the distinguishability between diverse image categories.The combination of CNNs and Transformers not only extracts rich local and global information but also captures unique features of images on a point-to-point basis, thereby improving classification accuracy. Experiments were conducted on the Mini-ImageNet, CIFAR100, and CIFAR10 public datasets. The results show that ARNet achieved top-1 accuracies of 65.21%, 78.32%, and 93.39% on Mini-ImageNet, CIFAR100, and CIFAR10, respectively, outperforming other models such as RMT, TCFormer, and SSViT. Additionally, we applied ARNet at the Shandong base of the national wood industry engineering research center to identify transverse section micrographs of 20 precious wood types, achieving an accuracy of 99.50% on the test set. After loading the parameters into the re-parameterized model, the validation set accuracy was 99.20%, with a detection time of 0.024s per image. This demonstrates that by combining residual dynamic convolution with multi-scale dilated attention, the accuracy and real-time performance of wood micrograph classification can be effectively improved.