Research Of Chip Character Detection Based On Domain Adaptation

Character recognition is a valued technology in academia and industry.From vehicle license plate recognition to text translation,character recognition can be used to solve related problems based on targeted scene settings and model design.Chip character recognition,a special scene of character recognition,can be uesd to solve industrial problems such as industrial defect detection and automatic chip assembly.In early years,chip character recognition methods such as template matching could only work in fixed fonts and fixed scenes.While,in recent years,with the expansion of deep learning algorithms and the increase in floating-point performance of graphics cards,deep learning models can recognize more similar fonts and more scenes of chips.Apart from this,the high precision of deep learning models is based on large and extensive character image data sets.However,due to commercial restrictions on the chip character data set,there are fewer public character pictures and the collected pictures are also quite vague,which can hinder the training of the model.In response to the above problems,this article aims to establish a domain-adaptive chip character recognition model.The main contents of this article are as follows:(1)In order to deal with the small sample learning where the public character dataset is insufficient to meet the demand of the common character position detection model,this paper integrates the adversarial learning method into the chip character position detection algorithm,which can improve the detection accuracy of the chip character position detection algorithm under a few samples.The algorithm makes the source domain and target domain data close to each other at the feature space,and completes the domain adaptive optimization based on the labeled samples of the source domain and the unlabeled samples of the target domain..Meanwhile,in order to achieve a better effect of the algorithm on industrial chip images,this paper also proposes a targeted solution for image segmentation.(2)This paper establishes a character recognition model,which can achieve acceptable recognition effect even when the target data set is hard to label and the number is small.Based on the domain discriminator and the gradient flip layer that measure the distance between domains,as well as the discrimination method of weight matrix attention mechanism and the output of the classifier,this model layers the multidimensional feature images into different domain discriminators to reduce the phenomenon of negative transfer and enhance the recognition accuracy.(3)In addition to realizing the domain-adaptive chip character recognition algorithm,this paper also designs and develops a corresponding system based on it,and tests the system on the chip character dataset.The system solves the problem of defect detection when the image collected on the industrial production line is blurred and the production line is put into use for a short time,reducing the labor loss of manual chip sorting.