Research On Automatic Synthesis Algorithm And Detection Model Of Carton Dataset For Domain Generalization

Object detection technologies based on deep learning are widely used in logistics industry such as goods classification and detection,which require a large number of highquality labeling datasets to achieve good performance.At present,cartons have become the mainstream products in the logistics industry due to their good structural performance and low cost.However,there are some problems in the construction of the carton dataset,such as difficult data cleaning,long time-consuming image annotation,and difference(domain bias)between the constructed carton dataset and the data distribution of the task scene.Therefore,aiming at the above problems,this thesis focuses on the automatic synthesis algorithm of carton dataset in the target domain and the domain-adaptive carton detection model.The main research contents and contributions are as follows:1)According to the similar palletizing forms of cartons in different scenarios,this thesis proposes a texture decoupling algorithm for carton instances,which extracts the contour combination information and stacking structure information of different sides of the carton to save the information such as truncation,occlusion,scale transformation,and perspective transformation during imaging of the carton instance.And the carton contour skeleton is constructed,which can be used for the algorithm research of carton data automatic generation.2)Based on the characteristic that most complete contours of cartons are parallelograms or trapezoids,an incomplete contour reconstruction algorithm is proposed to construct an outer enveloping parallelogram with the smallest area for the incomplete carton instance.The algorithm makes the contextual semantic relations such as occlusion and truncation of the synthesized carton instance in the background more realistic.3)The automatic image synthesis algorithm with carton texture decoupling combination is proposed for building a target carton training dataset.By replacing the carton pattern,color,and other textures in the contour skeleton dataset with the carton texture of the target domain,the carton images in the target domain are synthesized.The domain bias between training datasets and the target task is alleviated and the deployment time of the carton detection system is shortened.4)Because the current domain adaptation theory ignores the coexistence of multiple features types,which make it difficult to improve the performance of the carton domain adaptation detection model,this thesis proposes a mixed-class feature distribution metric theory.When there are multiple class targets in the feature receptive field at the same time,it can be regarded as a mixed class to calculate the mixed class feature distribution distance.5)A semantic consistency domain adaptive detection model is proposed on basis of the distribution metric theory of mixed-class features,which achieves the semantic separability of single-class and mixed-class features through a semantic prediction network and a semantic bridging component,and utilizes the semantic attention domain loss to alleviate the imbalance of positive and negative samples.The experimental results show that the model can improve the adaptability of the detection model to the domain bias and improve the detection accuracy of the carton detection system in the task scene.6)The automatic synthesis algorithm of carton dataset and the domain adaptive carton detection model proposed in this thesis are tested.The results show that the automatic synthesis algorithm of carton dataset can effectively reduce the influence of the domain bias between the training dataset and target task,and the mAP of the single-stage model is improved by 4.3%～7.5% and the mAP of the two-stage model is improved by 3.4%～9.3%.The experiments on carton dataset and public dataset show that the semantically consistent domain adaptive detection model can effectively improve the adaptability of the detection model to the domain bias,it achieves mAP 6.7～7.7% improvement in carton dataset.The above algorithm and model have been applied and verified in the carton detection task of the intelligent loading and unloading mobile robot.And the grasping detection and positioning accuracy of the carton detection system is improved by 8%～9.4%.