Research On Intelligent Detection Technology Of Delivery Safety Based On Machine Vision

With the appearance and development of the Internet and e-commerce,people rely on online shopping and express delivery more than before.At the same time,the daily delivery volume of express logistics is also increasing steadily year by year,and the consequent delivery safety has become an unavoidable problem.With the rapid development and wide application of machine vision technology based on deep learning,it has become an important trend in the future to apply machine vision technology to complete automatic prohibited items detection in the delivery field.This paper focuses on the detection of prohibited items in X-ray images by using deep learning techniques.There are some difficulties in detecting prohibited items in X-ray images,for example,it is difficult to collect positive sample data,complex image background and occlusion,the large shape difference caused by multiple viewing angles of positive targets.In order to solve the above problems,this paper conducts relevant exploration and research on the object detector model from the aspects of transfer learning,attention mechanism,loss function,knowledge distillation and few-shot learning and so on.The main contributions of this paper include the following aspects.(1)In view of the difficulty in collecting positive sample data of Xray images and the lack of sufficient datasets to train complex models,this paper proposes a shallow transfer learning method based on fine-tuning,which uses a specific "incremental network structure+partial fine-tuning"method to achieve transfer learning.Compared with the traditional finetuning,the performance of this algorithm is better.This method can quickly converge the complex model trained on the smaller X-ray image data set,and transfers the detection "ability" of SSD300 from natural light image datasets to the X-ray image datasets.(2)In view of the complex background of the sealed package in the X-ray image and the lack of color visual features in X-ray images,this paper proposes the attention mechanism of channel context block to improve the robustness of backbone network of target detection model.This paper uses a heat map to demonstrate the structure’s ability to "focus"on the target on the image through visual analysis experiments.Finally,experiments on prohibited items detection show that applying this attention mechanism structure to common object detectors can effectively improve the performance of prohibited items detection in X-ray images.(3)In view of the large shape difference caused by multiple viewing angles of positive targets,this paper proposes a truncated loss function.This paper demonstrates that this truncated loss function can force classification models to widen the cosine gap between features extracted from different classes.The experiments on CIFAR-10 prove that this truncated loss function can effectively improve the classification performance of the model.The experiments on prohibited items detection show that applying this truncated loss function to the object detection model can effectively improve the detection performance of prohibited items in X-ray images.(4)As for the difficulty in detecting prohibited items from overlapped items of sealed package in X-ray images,this paper proposes the use of cooperative knowledge distillation algorithm.The research uses the teacher model to mine image-level and instance-level difficult samples,and enhances the performance of the student model.Different from the knowledge distillation method in which student imitates the teacher,this algorithm is realized by the cooperation of teacher and student.(5)This paper completes the research and development of an intelligent detection system for prohibited items based on X-ray security inspection machine.On the one hand,this paper proposes a two-stage fewshot object detection algorithm based on an incremental structure.It is verified by experiments that this method can not only effectively expand the detection of new categories of prohibited items with only a small number of samples,but also effectively maintain the detection ability that the object detector model learned in the underlying dataset.On the other hand,this paper designed and developed a real-time online prohibited items intelligent detection system based on distributed development.