Attention-based Image Compression in Sensor Assembly

The rapid development of the Industrial Internet of Things (IIoT) has led to an explosion in the amount of data to be transmitted and stored in the industrial sector. The investigated images are used in particular for quality control after individual process steps as part of Automated Optical Inspections (AOI). Image processing systems are among the key technologies here. Nevertheless, the storage and transmission of image data is still a major challenge for many companies. The use of AI-based image compression methods with higher quality performance compared to conventional methods becomes inevitable. Convolutional Neural Networks (CNNs) or Transformers are currently gaining ground in this area. However, the efficiency of these architectures has not yet been demonstrated in an industrial scenario. In this work, both a CNN- and a Transformer-based model with built-in Attention Mechanisms are trained with 8 bits per pixel (BPP) grayscale images generated in an industrial ultrasonic sensor production line. Even in the first test phase, in which only 2,000 images were used for model training, a compression rate of 0.162 BPP is achieved, far exceeding the quality of comparable JPEG-compressed images. When training the models with 100,000 images, the lowest BPP rate of 0.009 is achieved by the Transformer-based algorithm, resulting in an image quality that is sufficient for most quality assessment use cases. Thus, our results demonstrate the applicability and superiority of AI-based compression algorithms in industry. Their use leads not only to more efficient data transmission, but also to huge savings in storage space and costs for companies compared to traditional image compression methods.