Research On Failure Patterns And Metamorphic Testing Techniques For Object Detection Networks

Object detection is one of the important application directions of deep learning technology.In safety-critical fields such as autonomous driving and security monitoring,a failure of the object detection network due to its own quality may lead to significant loss of life and property safety.Therefore,the reliability of the object detection network has become one of the important prerequisites for expanding the application scope of object detection technology.Deep learning-based object detection networks have fundamentally different characteristics in terms of architecture and working logic compared to traditional software: on the one hand,the implementation code of object detection networks is relatively simple but highly dependent on underlying deep learning frameworks;on the other hand,the recognition performance of object detection networks is highly dependent on high-quality training data.Therefore,the reliability assurance work of object detection networks has significant differences from that of traditional software reliability assurance work.To achieve the research goal of reliability assurance of object detection networks,this paper conducts empirical study on failure patterns in the implementation program of object detection networks,summarizes the root causes of failure patterns,and gives suggestions.Based on this,further research on testing techniques for object detection networks is conducted,and a metamorphic testing technique for object detection networks is proposed.Test data is generated based on metamorphic relations,and the test results are judged.The main works are as follows:(1)Empirical study on failure patterns of object detection networks.Conducts empirical study on 101 programs implemented by 104 students in a software development course for object detection tasks,summarizes 13 common failure patterns,analyzes six root causes of object detection network failures,and gives suggestions for the students and entry-level developers to avoid potential risks due to these failures.The study finds that 45.81% of the failures are due to code-related issues,32.26% are caused by incompatible execution environments,17.74% are specific to object detection networks,and 4.19% are due to errors in training data.This prompt serves as a reminder for developers to pay attention to issues such as the compatibility of deep learning frameworks and adherence to code standards when implementing object detection networks.(2)Research on metamorphic testing techniques for object detection networks.In response to the difficulties in constructing test data and judging test results in the testing of object detection networks,a metamorphic testing technique called OBMeta is proposed for object detection networks.This technique separates images into foreground and background,and uses four types of graphic transformation methods to construct diversified metamorphic relations for the foreground,background,and the entire image.Based on these metamorphic relations,OBMeta extends and generates derived test data and matching test expected results from the original test data.To further improve the quality of the test data,OBMeta also uses image similarity algorithms to screen out obviously unreasonable image data with low similarity to the real image.(3)To validate the effectiveness of the technique,experiments were conducted on six object detection networks trained on the MS COCO2017 dataset.The experimental results demonstrate that,compared with existing metamorphic testing techniques,OBMeta,which uses more diversified metamorphic relations,can construct more diverse test data with lower time costs.These test data,compared with those generated by existing metamorphic testing techniques,can produce more data with error detection capabilities and have higher similarity to real images.Furthermore,by using adversarial examples to retrain the object detection networks,the robustness of the object detection networks can be further improved.These results indicate the effectiveness of OBMeta in enhancing the reliability of object detection networks.