毕业生知识库

随着生活水平和消费能力的提高，人们购买越来越多的液体商品，例如饮料、乳制品、酒等，这极大地刺激了企业生产。随着液体商品产量的增长，瓶子的需求量也在逐年迅猛增长。同时，消费者高度关注食品安全。而对于液体商品生产企业，要肩负起主体责任，在进行液体商品灌装前，对空瓶进行缺陷检测，以提供符合行业标准操作程序的高质量产品。目前，国内外针对空瓶的缺陷检测大多聚焦于透明瓶子瓶底、瓶口、瓶身的缺陷检测和不透明瓶外观的缺陷检测。而由于不透明空瓶内部检测的难度较大，几乎没有针对不透明空瓶内壁的缺陷检测算法及机器。

基于此背景，本文在组内前期工作的基础上，设计了一套不透明空瓶内壁缺陷检测装置，并构建了相应的缺陷数据集。针对该场景，本文有针对性地提出了基于数据增广的不透明空瓶内壁缺陷检测的方法，并对检测模型的轻量化进行了研究。本文主要研究内容如下：

（1）基于USB3.0的分布式图像快速采集系统

根据前期工作存在的诸多问题，本文设计了一套图像采集系统，该系统设计开发基于 Cypress公司的 EZ-USB芯片CYUSB3065的USB3.0外设解决方案。该系统采用CMOS图像传感器IMX335LQN-C进行图像采集，并利用USB3.0传输方式实现了CMOS图像传感器与PC的高速数据传输。该系统在结构上具有体积小，便捷性高等特点，在采集性能上具有可靠性好，图像质量高等特点。此外，本文基于该设备构建包含各类缺陷类型的不透明空瓶内壁缺陷检测数据集。

（2）基于数据增广的YOLOv5不透明空瓶内壁缺陷检测模型

受训练数据不平衡的影响，缺陷检测模型往往存在鲁棒性差等问题。为进一步提升缺陷检测模型的性能，本文提出了一种基于递归数据增广框架的YOLOv5缺陷检测模型。该框架能够根据已有数据的特点在缺陷块源库构成与数据平衡策略方面进行灵活调整，从而在不改动YOLOv5模型结构的基础上提升模型的缺陷检测能力。丰富的框架对比实验与模型性能对比实验验证了基于该数据增广框架的YOLOv5模型在不透明空瓶内壁缺陷检测数据集上的良好性能表现。

（3）面向不透明空瓶内壁缺陷检测的轻量化方法研究

在实际生产过程中，缺陷检测模型常存在算力要求高、实时性差等问题。这些问题会导致性能较高的大尺寸模型无法进行有效部署，但同时参数量过小的模型在性能上存在瓶颈，无法实现准确的缺陷检测。针对这些问题，本文构建了一种面向不透明空瓶内壁缺陷的轻量化合作检测模型。该合作模型包含一个预训练的高性能教师模型与一个经知识蒸馏微调的高效率学生模型。丰富的定量与定性对比实验表明该合作模型可在识别性能与识别速度方面达到高效的平衡，在本文的缺陷数据集上具有良好的检测性能。

With the improvement of living standards and consumption capacity, people buy more and more liquid goods, such as beverages, dairy products, wine, etc., which greatly stimulates the production of enterprises. With the growth of the output of liquid commodities, the demand for bottles is also growing rapidly year by year. At the same time, consumers pay high attention to food safety. As for liquid commodity manufacturers, they should shoulder the main responsibility to detect the defects of empty bottles before filling liquid commodities, so as to provide high-quality products that meet the industry standard operating procedures. At present, the defect detection of empty bottles at home and abroad mostly focuses on the defect detection of the bottom, mouth and body of transparent bottles and the appearance of opaque bottles. However, due to the difficulty of internal detection of opaque empty bottles, there are few defect detection algorithms and machines for the inner wall of opaque empty bottles.

Based on this background, the thesis designs a set of defect detection device for the inner wall of the opaque empty bottle and constructs the corresponding defect data set on the basis of the previous work in the group. Aiming at this scenario, the thesis proposes a method for detecting the defects on the inner wall of the opaque empty bottle based on data augmentation, and includes the research on the lightweight of the detection model. The main research contents of the thesis are as follows:

(1) Distributed image fast acquisition system based on USB3.0

According to many problems existing in the previous work, the thesis designs an image acquisition system, which is based on the USB3.0 peripheral solution of Cypress's EZ-USB chip CYUSB3065. The system uses CMOS image sensor IMX335LQN-C for image acquisition, and uses USB3.0 transmission mode to realize high-speed data transmission between CMOS image sensor and PC. The system has the characteristics of small size, high convenience, good reliability and high image quality in terms of acquisition performance. In addition, based on the equipment, the thesis constructs a data set for the detection of the defects on the inner wall of the opaque empty bottle. The data set contains various types of defects.

(2) YOLOv5 detection model for defects on the inner wall of the opaque empty bottle based on data augmentation

Affected by the imbalance of training data, defect detection models often have problems such as poor robustness. To further improve the performance of defect detection model, the thesis proposes a YOLOv5 defect detection model based on recursive data augmentation framework. The framework can flexibly adjust the composition of defect block source library and data balance strategy according to the characteristics of existing data, so as to improve the defect detection capability of the model without changing the structure of YOLOv5 model. Abundant frame comparison experiments and model performance comparison experiments have verified the good performance of YOLOv5 model based on the data augmentation framework on the data set for the detection of the defects on the inner wall of the opaque empty bottle.

(3) Research on lightweight method for detecting defects on the inner wall of the opaque empty bottle

In the actual production process, the defect detection model often has problems such as high computing power requirements and poor real-time performance. These problems will cause the large-scale model with high performance to be unable to be deployed effectively, but at the same time, the model with too small parameters has a bottleneck in performance and cannot achieve accurate defect detection. To solve these problems, the thesis constructs a lightweight cooperative detection model for the detection of defects on the inner wall of the opaque empty bottle. The cooperation model includes a pre-trained high-performance teacher model and an efficient student model finely tuned by knowledge distillation. A wealth of quantitative and qualitative comparative experiments show that the cooperation model can achieve an efficient balance between recognition performance and recognition speed, and has good detection performance on the defect data set in the thesis.