KAM-Net: Kilobyte-Scale Ultralightweight Attention-Based Network for Glass Defect Detection With Algorithm/Hardware Co-Design

Glass manufacturing processes often yield complex defects such as bubbles, stones, and scratches. While machine vision systems powered by deep learning offer effective defect detection, real-time inference on production lines remains a hurdle. To address this, we propose KAM-Net, an ultralightweight kilobyte-scale defect detection model utilizing depthwise separable convolution (DSC) and improved triplet attention (TA). With just 4.2 kB of parameter storage, KAM-Net achieves impressive performance, achieving a 92.3% mIOU and $91.5\%~F1$ -score on the glass defect dataset, surpassing traditional models like UNet. It also demonstrates competitive performance in public classification datasets. Then, we design an efficient algorithm/hardware codesign for KAM-Net on field programmable gate array (FPGA). Leveraging parallelism and pipeline techniques, KAM-Net achieves a remarkable processing speed of 2564 FPS, outperforming GPU-based processors. Notably, our work marks the first hardware implementation of the TA mechanism, showcasing its practical application potential. Our detection system ensures high-precision defect identification, matching real-time image acquisition speeds of production lines.