This research designed and developed a deep-learning-based anomaly detection framework to detect core failures from die casting X-ray images. The area of interest can be located automatically. The recognition and the location of the defects were achieved using only one neural network. The noises introduced during the dimension reduction process were solved with an edge-detection-based generalizable noise removal approach. Our work achieved an impressive accuracy of 97.45% with only a tiny data set of 30 known-good images. Our proposed CAE-based deep learning framework is robust during the training and the testing stages. It is easy to develop and requires only a small amount of labeled data to train. The anomaly detection and location approach presented in this manuscript does not apply to the general object detection tasks. It is mostly suited to discover differences among similar images, such as the X-ray images of the same parts. Despite the fact that the framework is only tested on die casting images, the approach itself has a great potential to be applicable to other industrial inspection scenarios. There are several directions for future research on this topic. Firstly, while the proposed method is designed to be generalizable to inspect other core defects, it is essential to test the method on more scenarios. We believe our method can be adapted to other core inspection tasks effortlessly as discussed in Sect. 5. However, its applicability on more blurring defects, such as minor cracks or porosity, remains unrevealed. Secondly, GAN-based anomaly detection applied in surface defect detection [10,18] is still a possible route for core defects inspection. Although our preliminary trials of GAN produced unsatisfactory results, additional researches should be investigated to fully understand the potentials of GAN-based anomaly detection techniques. Finally, another fascinating research direction is to apply clustering algorithms to classify the inspected defects into different categories without any annotation. This will be beneficial for experienced engineers to figure out the problems in the system.
n this work, we introduce a Denser Feature Network (DenserNet) for visual localization. Our work provides three principal contributions. First, we develop a convolutional neural network (CNN) architecture which aggregates feature maps at different …
Aerial imagery has been increasingly adopted in mission-critical tasks, such as traffic surveillance, smart cities,and disaster assistance. However, identifying objects from aerial images faces the following challenges: 1) objects of …
In this paper, we aimed at implementing an application in detecting fire and other critical ground-based objects in a wildfire event using high resolution aerial images. We propose a well annotated fire dataset with 1400 4K images. We also present a coarse-to-fine strategy to deal with the 4K images, which achieves high accuracy while maintaining fast speeds. Our methods can also be added to different backbones in object detection methods and extended to deal with high resolution images. Ongoing and future research objectives involve expansion of the UAS wildfire imagery collection, and working with a UAS platforms equipped with more powerful CPUs and GPUs. Fusing data collected from multiple types of sensors can provide additional wisdom in wildfire fighting scenarios. Additional Machine Learning approaches, especially a hybrid approach that combines signal processing with deep learning, will be investigated to discover a faster and more accurate technique to identify small objects of interests and objects with irregular boundaries in high definition videos and images.
Intrusion detection is a pivotal step for network protection. Usually, intrusion detection is performed at packet level by using deep packet or state-full protocol inspection to detect malicious requests in the network. However, flow based analyses …
This project aims at applying machine learning to solve real-world use-inspired research questions.