中国科学院自动化研究所机构知识库

基于视觉感知的表面缺陷检测，具有高效、可防止二次损伤等优点，被广泛应用于工业质检领域。基于数据驱动的深度学习技术的快速发展进一步推动了视觉算法的进步与应用。然而，大规模工业图像的收集与标注在工业现场难度较大，工业图像与自然场景图像的巨大差异也导致一些通用的视觉算法不能直接应用。这些问题阻碍了深度学习方法在工业领域的规模化应用。为了解决这些问题，本文对表面缺陷检测任务中的常规监督、弱监督和异常检测算法展开深入研究，提出了适用于工业图像的一系列高效缺陷识别算法，以递进的方式降低工业数据的采集与标注成本。本文所提算法可满足不同硬件条件、数据形式的需求，在工业领域具有较强的实用性。本文的主要工作和贡献如下：
1.针对工业图像的高分辨率和纹理信息丰富的特性，本文提出了一种工业场景下的高效表面缺陷检测算法。该算法以常规监督的形式，采用轻量化的骨干网络，并且提出了一种有偏特征融合网络，加强了对于图像底层纹理特征的利用，有效地提升了检测精度。另外，该算法设计了一种全新的联合缩放策略，通过单一参数对网络结构进行优化，得到一系列可适应不同硬件约束的缺陷检测模型。实验表明，该算法相比现有的单阶段目标检测算法，缺陷检测精度更高，计算量和参数量更低。
2.针对正矩形检测框背景区域过大、无法有效提取目标形态信息的问题，本文提出了一种面向工业图像中不规则目标的四边形检测算法。该算法首先设计了一组八维坐标表征被测目标的外接四边形，然后对经典目标检测算法中的区域建议网络、区域池化操作等部件进行改进，以适应不规则四边形操作，使算法在常规监督形式下可以鲁棒地、准确地输出工业图像中检测目标的外接四边形。实验证明，该算法的检测结果对比常规的正矩形框包含了更多目标的形态信息，摒除无用背景信息的干扰，提高了后续进行工业产品质量判断的精度。
3.为了降低表面缺陷检测任务中的人工数据标注成本，本文提出了一种基于弱监督学习的缺陷目标检测算法。该算法只需具有图像标签的数据进行训练，可实现缺陷的目标检测任务，从而降低了在缺陷目标检测任务中手工标注缺陷包络框的成本。在该算法中，设计了弱监督卷积池化模块提取分类网络的特征图中的空间信息，并采用了知识蒸馏策略，在不降低实时性的前提下提升模型精度。同时，本文进一步探索了基于弱监督学习的实例分割算法的研究，结合人类视觉机制，利用便捷的人工涂鸦标注，实现弱监督条件下的实例分割，并在开源数据集上进行了验证。
4.为了进一步降低表面缺陷检测任务中的缺陷样本收集成本，本文提出了一种基于异常检测的缺陷图像识别算法。在该算法中，设计了包含一个编码器和一个生成对抗网络的背景重构网络，只需利用实际采集的正常样本进行训练，实现对于被测图像的背景重构，进而通过被测图像和重构图像的差异识别缺陷图像。通过在开源表面缺陷数据集的实际测评，该算法达到了更高的识别准确率。

With the advantages of high efficiency and preventing sencondary injury, inspection of surface defects based on visual perception has been widely used in industrial quality inspection field. The rapid development of data driven-based deep learning technology contributes further to the progress and the application of vision algorithms. However, it is difficult to collect and annotate large-scale industrial images in industrial sites. The great difference between industrial image and natural scene image also makes the existing general visual algorithms cannot be applied directly. These problems hinder the wide application of deep learning methods in the industrial field. To solve these problems, this dissertation carries out research on regularly supervised learning-based, weakly supervised learning-based and anomaly detection-based algorithms in the inspection task of surface defect. A series of efficient defect recognition algorithms for industrial images have been proposed, which gradually reduce the cost of industrial image collection and annotation. The algorithms proposed by this dissertation can overcome the limitations of various hardware conditions and data forms, and have strong practicability in the industrial field. The main contents and contributions of this dissertation are summarized as follow:

1. Considering the industrial image's characters of high resolution and rich texture information, this dissertation proposes an efficient surface defect detection algorithm in industrial scene. Based on regularly supervised learning, this algorithm adopts lightweight backbone and proposes a biased feature fusion network that can utilize more low-level texture features, so that the detection accuracy is improved effectively. Besides, this algorithm designs a new compound scaling strategy that can use only one hyper-parameter to jointly optimize the network architecture so that a family of defect detectors can be developed. As experimentally demonstrated in experiments that compared with the recent state-of-the-art detector, this algorithm achieves higher defect detection accuracy but with fewer parameters and lower computation.

2. Considering the problem that rectangular box includes much useless background information and could not provide efficient shape and posture information, this dissertation proposes quadrangular object detection for irregular target in industrial image. Firstly, an eight-dimensional vector is designed to represent the quadrangular bounding box of object. Then, the region proposal network, RoI pooling operation and other components in classical object detection algorithm are improved to adapt to the operation of quadrangular bounding box. So that this algorithm can robustly and accurately achieve quadrangular output. It is demonstrated in experiments that compared with the rectangular box, quadrangular output of this algorithm can provide more shape and posture information for object and eliminate the background noise, which improves the accuracy of the following industrial quality inspection.

3. In order to reduce the cost of manual data annotation in surface defect detection task, this dissertation proposes a weakly supervised learning-based defect detection algorithm. This algorithm can achieve the defect object detection task with only image-tag annotation for network training, which reduces the cost of manual annotation of defect bounding box in defect detection task. In this algorithm, a weakly supervised learning-based Conv-Pooling module is designed to extract spatial information from the feature map of classification network. The knowledge distillation strategy is also adopted to improve the accuracy while without degrading the real-time inference. Meanwhile, this dissertation further explores the research on weakly supervised learning-based instance segmentation. The proposed algorithm borrows from the process of human perception and utilizes practical scribble annotation to achieve the instance segmentation. Experiments conducted on open-source dataset verify the effectiveness.

4. In order to further reduce the cost of defect image collection in surface defect inspection task, this dissertation proposes an anomaly detection-based defect image recognition algorithm. In this algorithm, a background reconstruction network that consists of an encoder and a generative adversarial network is designed. This network can achieve background reconstruction for testing image with only normal image for network training, so that defect image recognition can be finished by comparing the testing image and is reconstructed image. Experiments conducted on open-source defect inspection dataset verify that this algorithm achieves higher accuracy than the previous methods.