关键词: 羽绒, 羽绒图像分类, 动态聚焦网络, 注意力机制, 动态间隙关注损失函数, 三阶段级联训练, 残差网络

Abstract:

Objective In down quality evaluation systems, key indicators such as color and freshness are decisive factors for product grading and market value. For fine-grained quality classification tasks involving down characteristics like color and freshness, traditional manual inspection methods exhibit low efficiency and strong subjectivity, while existing computational approaches demonstrate inadequate capability in recognizing the intricate texture patterns of down materials. To battle these challenges, this study proposes a Tri-phase residual dynamic focus network (Tri-RDFNet) to enhance the classification accuracy of down images, thereby advancing automated quality assessment in the down industry.

Method Building upon an enhanced ResNet architecture, a novel three-stage residual dynamic focus network was developed for fine-grained classification of down feather images. The network incorporated dilated convolution modules, deformable spatial-convolutional block attention module (DS-CBAM), and dynamic gap-aware attention loss (DGALoss) function to enable in-depth feature learning of down images. Furthermore, a three-stage cascaded training strategy was introduced to significantly improve the model's generalization capability. The experimental dataset, collected using industrial camera systems with a bar-shaped white light source, comprised four categories, i.e. white fresh down, white recycled down, colored fresh down, and colored recycled down. Comprehensive experiments were carried out using the established model.

Result Comparative experimental results demonstrated that traditional CNN models, such as AlexNet and VGG, exhibited certain performance bottlenecks in this task, achieving accuracy rates of only 89.32% and 90.26%, respectively. These models struggled to capture the fine-grained differences inherent in down feather images. Although Transformer-based models possess strong global modeling capabilities, they suffer from overfitting due to the limited dataset size and architectural complexity. The backbone model RDFNet enhanced the learning focus on down image features by incorporating atrous convolution modules, DS-CBAM, and DGALoss. As a result, it achieved a classification accuracy of 95.28% on the collected down image dataset, representing an improvement of 1.11%-5.96% compared to traditional models such as AlexNet, VGG, ResNet, ViT, and Swin Transformer. Furthermore, based on this RDFNet backbone, a three-stage cascaded training strategy was introduced. In the first stage, the model was trained globally using the cross-entropy loss function. The second stage employed DGALoss to reweight and train on easily confused samples, yielding a 1.02% increase in accuracy over the first stage. In the third stage, noise samples were filtered, and sample weights were reassigned to further train the retained samples. This final phase enhanced the model accuracy by an additional 0.71%, achieving a final accuracy of 97.01%. This three-stage process reduced the risk of overfitting while improving precision and generalization. Ablation studies confirmed the effectiveness of each component. The atrous convolution module improved the model's ability to perceive multi-scale features in down images, raising the validation accuracy by 0.99%. The DS-CBAM module enhanced the model's feature selection capability by integrating channel attention with deformable spatial convolution, leading to further improvement in accuracy while introducing minor overfitting. When combined, DS-CBAM and atrous convolution boosted accuracy to 95.28%. Introducing the three-stage training scheme and applying FocalLoss during the second stage to focus on hard examples increased accuracy to 96.11%, thus improving model robustness and stability. Replacing FocalLoss with DGALoss for better focus on confusing samples led to the highest validation accuracy of 97.01%, demonstrating DGALoss's superior capability in distinguishing ambiguous down categories.

Conclusion To address the challenge of fine-grained classification in down feather images, this paper proposes an innovative three-stage residual dynamic focusing network. The core backbone model, RDFNet, enhances feature extraction capabilities by improving the ResNet architecture through the integration of DS-CBAM and atrous convolution modules. Based on RDFNet, a three-stage training strategy is designed, consisting of warm-up training, adaptive weighted training using the novel DGALoss function, and refined sample training, collectively forming the Tri-RDFNet model. This approach effectively improves the recognition of easily confused down feather image samples and enhances the model's generalization ability. Experimental results demonstrate that the proposed method achieves a classification accuracy of 97.01% on a self-constructed dataset of 8 000 down feather images, significantly outperforming traditional methods. This provides an efficient solution for automated down quality assessment and offers a valuable reference for fine-grained image classification tasks.

Key words: down, down image classification, dynamic focusing network, attention mechanism, DGALoss, Tri-stage cascaded training, residual network