DETRs with Dynamic Contrastive Denoising Training for Smartphone Assembly Parts

DETRs with Dynamic Contrastive Denoising Training for Smartphone Assembly Parts

In the scenario of 3C (Computer, Communication, Consumer Electronics), the algorithm for detecting targets in smartphone component assembly consumes a substantial amount of system computing resources.It also faces challenges such as the flexible nature of target components and the small scale of heterogeneous components, leading to low detection accuracy. To adapt to the 3C scenario, this paper proposes improvements based on the DINO object detection model. It introduces a more lightweight and powerful feature extraction backbone, Efficientnetv2, and utilizes the He-Kaiming weight initialization method to extract strong multi-scale feature maps. In training, a more efficient dynamic contrastive denoising training method is employed. This approach makes the model lightweight and accurate for 3C detection. This method outperforms leading detection algorithms in both accuracy of experimental results and parameter efficiency.