Object Detection Of Intelligent Container Based On Human-machine Cooperation And Self-supervised Learning

Recent years,benefiting from the success of deep convolutional neural networks,object detection has made progress in accuracy and efficiency.However,behind the progress of the algorithm,massive data and annotations are essential.Although a large number of natural images can be derived from cameras or mobile devices,the annotations used for training,need to determine the category labels and bounding boxes.The labor cost is high,especially in the scenes that require professional knowledge or complex labeling difficult to obtain.Although challenging,due to its importance,it is attracting increasing attention that how to use large-scale unlabeled or partially labeled data to reduce labeling costs and enhance model performance.As a typical application of object detection,because of the difference and diversity of product categories,the huge workload of labeling has seriously affected the progress of research and application in the field of intelligent retail.First,we introduces our proposed IRC dataset.The dataset covers more than 200 common retail products in 10 major categories,collected in a real unmanned freezer environment,including more than 30,000 pictures and 370,000 object annotations.Meanwhile,considering the complexity of the actual environment,we artificially added a variety of abnormal scenarios.Evaluations show that the dataset is extremely challenging and differences in categories.In order to solve the problem of labeling cost in large scale dataset,we propose a human-machine collaborative object detection framework SPAL.The framework uses only a small number of annotations to initialize the detector.During the learning process,active learning is used to mine unlabeled images,and valuable samples are selected for manual annotation.The idea of self-paced learning is used,from easy to difficult to give credible pseudo labels to the remaining samples.During the iteration process,the change of the selected parameters is controlled.Finally a well-labeled complete dataset and a better detection model are obtained.Comparative experiments on the PASCAL VOC and IRC datasets show that our proposed framework has good robustness and performs well on small data volumes.Compared with other continuous learning frameworks,SPAL can achieve comparable detection performance with fewer annotations,avoiding the problem of accuracy degradation caused by false annotation error accumulation.Further,in view of the fast-changing characteristics of the smart retail category and the domain prior knowledge of the IRC dataset,we propose a self-supervised incremental learning machine SILM.When adding a category not included in the original detector,without a new category label,construct a feature extractor that shares parameters with the detector,iteratively label samples and update the training set,get a model that adapts to the unlabeled new category.Experimental results show that the method can automatically label based on categories,which greatly reduces the workload and improves the rapid adaptability of the model.