面向资源受限农业边缘设备的轻量化视觉Transformer压缩框架

Summary of Translation Process

Okay, here's how I approached this translation. I started by breaking down the source text, an abstract about Vision Transformer (ViT) compression for chilli crop disease detection, into individual sentences. My goal was to produce a high-quality Simplified Chinese translation. I paid close attention to technical terms, ensuring their accurate representation in the target language.

I began by analyzing each sentence. For instance, phrases like "Vision Transformers (ViTs)" became 视觉Transformer (ViT), carefully preserving the acronym and adding the relevant Chinese description. Terms such as "resource constrained devices" were translated as 资源受限的设备, aiming for clarity and precision. I had to make decisions about how to best represent concepts like "out-of-distribution test split", which I rendered as 分布外（OOD）测试集划分 to maintain the original meaning.

Next, I drafted a translation for each sentence, striving for both accuracy and natural flow in Simplified Chinese. I considered different phrasing options for terms like "economically significant crops," settling on 最具经济价值的作物之一 to capture the nuance. I ensured the terms "pruning," "quantization," and "knowledge distillation" were correctly translated as 剪枝、量化和知识蒸馏. When composing the whole sentences, I sought to capture the meaning of the combined compression steps, like in the translation of: "Hessian-Balanced Adaptive Block Pruning (H-BAC), guided by second-order sensitivity estimation, with quantization and attention-based knowledge distillation." I then went back and refined each sentence to enhance the overall cohesiveness and readability of the text.

I paid particular attention to the nuances of the compression techniques and the agricultural context. The term "sequential deployment pipeline" was adjusted to 顺序部署流水线, which is more appropriate in this context. The aim of identifying the "most effective configuration within each compression family" was translated as 为系统地确定各类压缩技术中最有效的配置.

After drafting the individual sentences, I carefully integrated them to ensure the translated text read fluently as a single cohesive unit. I checked the terminology to ensure uniformity and accuracy, and performed a final quality check for overall flow, accuracy, and professional tone. The final version maintains the original meaning while adhering to the structure specified by the prompt.

辣椒（Capsicum annuum）是印度最重要的经济作物之一，然而其产量长期受到病害的持续威胁，而这些病害在缺乏专家介入的情况下往往难以识别。尽管视觉Transformer（ViT）取得了极高的分类准确率，但其庞大的计算开销使得在资源受限设备上的部署极具挑战。现有的压缩方法通常将剪枝、量化和知识蒸馏割裂开来单独处理，导致对其联合应用的潜在收益与相互作用探索不足。为此，我们提出了一个统一的视觉Transformer压缩框架，将以二阶灵敏度估计为指导的Hessian平衡自适应块剪枝（H-BAC）与量化及基于注意力的知识蒸馏相结合。为了系统地确定各类压缩技术中最有效的配置，我们首先通过受控消融实验对各项技术进行独立评估，随后将表现最佳的组件整合到一个专为实际农业约束量身定制的顺序部署流水线中。在一个具有真实跨村庄、跨设备分布外测试集划分的辣椒三分类村庄划分数据集上，所得压缩模型的准确率达到或超过了95.13%的FP32基线水平，同时模型体积缩减了74%至98%；且在四种测试配置下，完全集成的压缩流水线在保持95.13 ± 2.32%准确率的同时，实现了54.5倍的模型体积缩减（从327.42 MB降至6.01 MB）。进一步的直接对比表明，在该数据集上，未经剪枝或蒸馏而直接训练出的相同最终尺寸学生模型，在同为6.01 MB的INT8大小下达到了相近的94.87%准确率，这揭示了H-BAC与知识蒸馏在何处值得付出其计算成本，而在何处其价值尚未得到体现。