毕业生知识库

禽类养殖业是国家经济的重要组成部分，随着养殖业的规模化发展，禽类疫病问题日益凸显，成为制约禽类养殖业健康发展的重大障碍。及时准确地诊断禽类疫病，采取有效的预防控制措施，对于保障禽类养殖业的健康发展至关重要。发展基于计算机视觉的疫病自动检测技术，是现代化疫病检测的新趋势。然而，一方面，以深度神经网络为基础的计算机视觉模型通常依赖于大量的训练数据以获得较高的识别精度，另一方面，受到国家疫病管控政策的严格限制，疫病的图像数据难以获得，这在很大程度上限制了深度学习模型的性能。为了解决以上问题，本研究旨在探索基于扩散模型的图像生成技术，实现禽类疫病图像数据增强，以提高下游相关机器学习模型的性能，进而提高禽类疫病的诊断效率和准确性。

为了实现禽类疫病图像数据的生成，本文沿着两条技术路线展开研究：第一，研究少样本图像生成技术。由于可获得的疫病图像数量往往很少，需要设计算法从有限的训练样本中学习到新概念的关键特征，从而准确地生成新概念的样本。利用该技术，可以生成特定类别的疫病样本图像或特定风格的健康样本图像。第二，研究基于指令编辑的疫病图像样本生成技术。根据疫病特征的文本描述和少量疫病参考图像对健康样本进行编辑，以平衡数据的分布，并在健康样本上生成疫病症状，进一步扩充疫病图像数据集。使用这两类方法生成相应的禽类的正常及疫病样本，为畜禽疫病的早期诊断机器学习模型的训练与优化提供了重要支持。本文的主要工作和贡献如下：

1. 提出一种基于特征对比增强的少样本图像生成方法。现有的少样本图像生成方法缺少显式的机制来捕捉训练样本中新概念所具有的关键特征，导致模型难以非常准确地生成新概念样本。针对这一问题，本研究在文生图模型的微调过程中引入了对比学习机制，通过将新概念样本与常规样本进行对比，使生成器更准确地捕捉到给定新概念样本的关键特征。在通用数据集WiKiArt、CUB和Oxford Flowers上的实验结果表明，该方法生成的图像能够更准确地生成新概念的样本。在禽类疫病数据集上的实验表明，该方法能够成功捕捉并生成具有症状特征的疫病样本，并能生成高质量的特定风格的健康样本。
2. 提出一种基于注意力权重特征融合的图像编辑方法，能够根据编辑指令执行图像编辑任务。针对现有指令编辑技术中存在的过度编辑问题，本研究提出一种即插即用且无需训练的方法，利用注意力图作为权重，在扩散模型的逆向过程中加权融合原图特征和去噪结果，提高编辑结果的区域准确性和内容连续性。在通用数据集IP2P和CWLD上的实验结果表明，该方法显著减轻了过度编辑现象；在禽类疫病数据集上的实验表明，一方面，该方法可以对现有疫病图像的风格进行编辑，获得更多样的数据，从而解决疫病数据分布不平衡的问题；另一方面，该方法还可以根据疫病特征的详细文本描述和疫病参考图像，在健康样本上生成包含疫病特征的新图像。与传统图像增强方法和其他基于图像编辑的数据增强方法相比，本方法对禽类疫病分类器性能的提升更为显著。

Poultry farming is an essential component of the national economy. With the expansion and digitization of the industry, poultry diseases have increasingly emerged as a significant barrier to its healthy growth. Timely and accurate diagnosis, along with effective preventive measures, is crucial for the sustainable development of poultry farming. The development of computer vision-based automatic disease detection is a modern trend. However, computer vision models based on deep neural networks generally require a large amount of training data for high accuracy. Stringent disease control policies often limit access to disease image data, thereby hindering deep learning model performance. To address these challenges, this study aims to explore image generation technologies based on diffusion models to enhance poultry disease image data and improve downstream machine learning models' performance, thereby increasing diagnostic efficiency and accuracy.

This study follows two technical paths to generate poultry disease image data. First, we investigate few-shot image generation. Given the limited availability of disease images, algorithms must be developed to capture the key features of new concepts from the available training samples, allowing for accurate generation of new concept samples. This technology can generate specific disease sample images or healthy samples in specific styles. Second, we study instruction-based disease sample generation. This approach edits healthy samples based on textual descriptions of disease characteristics and a few reference images to balance data distribution and introduce disease symptoms into healthy samples, further enriching the disease image dataset. These methods generate corresponding normal and diseased poultry samples, providing essential support for training and optimizing early diagnosis machine learning models for livestock diseases. The main contributions of this work are as follows:

1. We propose a feature contrast-enhanced few-shot image generation method. Existing few-shot generation techniques lack an explicit mechanism to capture the key features of new concepts in training samples, often leading to inaccuracies in generating new concept images. To address this, our research introduces a contrastive learning mechanism during fine-tuning of the text-to-image model, allowing the generator to accurately capture new concept features by comparing them to regular samples. Experiments on the WiKiArt, CUB, and Oxford Flowers datasets show that this method accurately generates new concept samples. Experiments on poultry disease datasets demonstrate that the method successfully captures and generates diseased samples with symptomatic features and high-quality healthy samples in specific styles.
2. We present an attention-weighted feature-fusion image editing method that allows image editing based on user instructions. To mitigate over-editing issues found in existing techniques, this plug-and-play, training-free approach uses attention maps as weights to integrate original image features and denoised results during the diffusion model's reverse process. This enhances the regional accuracy and continuity of the edited content. Experiments on the IP2P and CWLD datasets show a significant reduction in over-editing issues. On poultry disease datasets, the method not only allows style editing of existing disease images to create diverse data for balancing distribution but also generates new images with disease characteristics on healthy samples based on textual descriptions and reference images. Compared to traditional image augmentation and other image editing-based data enhancement methods, this approach significantly improves the performance of poultry disease classifiers.