Technology And Application Of Scrap Fine-grained Identification Under Complex Scenarios

China is the world’s largest producer and consumer of steel,and steel scrap is one of the important raw materials for the modern steel industry,as well as a renewable resource that is green and can be recycled many times.However,the traditional steel scrap recycling process requires quality inspection staff to carry out in a harsh environment slowly and quality is difficult to be guaranteed.To this end,this paper aims to design a high speed and high accuracy lightweight scrap identification model suitable for engineering deployment by combining several techniques such as data augmentation,contrastive learning,object detection and knowledge distillation.This study will not only reduce the work hazard factor of quality inspection staff,but also promote the safety,accuracy and efficiency of steel inspection work and the healthy development of the steel industry.To address the problems of small object area ratio,background clutter and detection difficulty due to the large resolution of the original scrap steel collection data,this paper proposes a set of data augmentation cropping strategy to solve the challenge of small object detection simply and effectively by increasing the object area ratio in both the training and testing phases.To address the difficulties of scrap data labeling and lack of public datasets,this paper proposes to construct powerful feature extraction networks by making full use of the information of unlabeled data itself through a comparative learning-based approach,and use it for scrap identification tasks by fine-tuning,reducing the dependence of the task on labeled data.To address the constraints faced by deep learning models in engineering deployments,this paper proposes to distill the knowledge from the teacher model with better performance but larger volume to the student model with smaller volume by adding non-maximum suppression and object confidence coefficients to knowledge distillation to obtain a lightweight scrap identification model with fast speed and high accuracy.In this paper,the YOLO v5 series,a single-stage object detection model that is currently popular in industry,is chosen as the base model:the YOLO v5x backbone network is used as the basis for the pre-training phase based on contrastive learning,YOLO v5x is used for the fine-tuning phase and is used as the subsequent teacher model,and the YOLO v5m is used as the student model for the distillation phase.To verify the effectiveness of this method in each phase,a large number of experiments were conducted on the validation and test sets.The experimental results show that the data augmentation cropping strategy proposed in this paper improves the detection accuracy of the student model by 4.6%;the detection accuracy of the method based on contrastive learning pre-training improves 4.7%and 6.6%on the teacher model and the student model respectively;the object detection knowledge distillation method can improve the model speed by 9.6fps and reduce the memory by 17.7M,with only 1.5%loss of accuracy.