Automated image anomaly detection using legacy camera infrastructure for visual inspection

Abstract

The automatic monitoring of industrial environments, the production process and the products themselves represents an important ally to allow a quick identification of an irregular situation. The visual inspection area develops computer vision techniques for such monitoring. However, computer vision solutions to automate these processes generally require the acquisition of new hardware systems that incur the need to install a new infrastructure of cameras and associated hardware, which has a high financial and operational cost. Often, the requirement to acquire and install a new image acquisition system delays or even prevents the adoption of computer vision in industrial environments. In addition, the adoption of machine learning technology usually requires the generation of large annotated databases, which makes the project even more expensive, and can make it unfeasible. Thus, computer vision systems that can exploit the installed data capture capacity have considerable potential for adoption by the industry. On the other hand, the legacy sensor infrastructure is typically composed of heterogeneous systems installed, with different devices and diversity of specifications such as spatial resolution, acquisition rate, color models and others. There are technical challenges that need to be addressed with research in order for the appropriate technology to be successfully developed.This project presents a solution to this problem by proposing the implementation of monitoring and quality control systems based on domain adaptation, considering only the legacy camera infrastructure of the industries. The objective is the development of technology that allows the implementation of computer vision solutions without requiring the incorporation of new hardware for image and video acquisition. Considering the diversity of heterogeneous systems installed, the strategy will be data normalization together with domain adaptation so that different data sources are standardized and used in a learning pipeline that, with this characteristic, will require less data without requiring additional hardware cost and operational. For the development of the detection system itself, we will use two methodologies: Design Thinking to create a user experience that is suitable for our target audience and relevant to their use cases, and Agile Development to guide the software development process based on feedback from real customers who use our functional prototypes. This proposal for PIPE 1 focuses on a specific case to demonstrate the technical feasibility of our idea. It is an application of computer vision for fault detection and analysis and predictive and prescriptive maintenance programming. This case aims to analyze images of train trucks for visual inspection. This application is under discussion with a customer in the rail logistics area, who needs to analyze the images to detect problems that indicate the need for maintenance of their trains. This company has truck image acquisition systems, but most of the analysis is still done by humans. Our proposal is to develop a system that receives the images produced by legacy sensors and performs the analysis automatically using computer vision and machine learning. (AU)