Backbone Selection and Timing Optimization for Real-Time Defect Detection on Fast Industrial Conveyors

Abstract

This work presents a canned-food defect-detection method using the EfficientDet model with four backbones (MobileNet, EfficientNet, Swin-T, and ConvNeXt). The dataset included 8046 images with a resolution of 512 × 512 pixels. The performance criteria in this study accuracy “mAP”, computational cost “FLOPs”, and Frames Per Second “FPS”. The lightweight backbone with EfficientDet achieves a mean Average Precision (mAP) of 90% with MobileNet, while EfficientNet achieves mAP accuracy of 92% and 94%. The heavy backbone for EfficientDet (Swin-T, and ConvNeXt) achieves mAP accuracy of 96% and 98%. The main contribution of this study is to optimize the speed of conveyor belts in industrial production lines to considers the limitations of hardware computational resources and detection delays. To ensure efficiency, the timing analysis of the conveyor belt speed problem is based on the principle of maximum speed versus constrains using a deterministic method. Margin timing achieves a belt speed of 5 m/s. To ensure quality, a probabilistic timing analysis method is used to achieve a conveyor belt speed of 0.9 m/s.