Research On Key Technologies Of Equipment Part Recognition Based On Deep Learning

At present,a large number of traditional machine learning methods and deep learning methods have been applied to industrial parts identification.Fine-grained image recognition is a subfield that has received continuous attention in image recognition and has performed well in many professional application scenarios.Since industrial parts have the characteristics of small differences between classes and large differences within classes,it is one of the effective methods to classify parts by fine-grained recognition.Although the fine-grained recognition networks with weak supervision can effectively use the label information,the computational amount and complexity of these networks are not a small challenge for practical implementation in the aviation industry.Based on the self-built fine-grained part recognition dataset,this paper integrates deep learning techniques into the model,compares multiple weakly supervised networks,optimizes the limitations of feature extraction in fine-grained models,and proposes a lightweight-based fine-grained part classification framework.The main work of this article is as follows:(1)According to the research background of this paper,the industrial parts dataset is established.The quality of the dataset will affect the performance of the model.Due to the rich variety of industrial parts,small differences between some parts and multi-sided similarities,etc.,the selection and shooting of parts are particularly important.In the process of establishing the part dataset,multi-angle sample images are taken for each part,and the data enhancement methods that meet the characteristics of the dataset are adopted to lay a foundation for subsequent research work.(2)For a wide variety of industrial parts,this paper proposes a fine-grained part recognition model.Firstly,a number of classical classification networks are compared and verified that the feature extraction backbone network Res Net50 is suitable for the part recognition task in this paper.After that,the experiment adopts the training techniques of pre-training and fine-tuning parameters,adds a regularization method while retaining the network structure of the feature extractor,which means that using the noise injection method to help the model extract key visual features by weighting before the final fully connected layer,so as to effectively improve the learning ability to discriminate features in images.(3)In order to adapt to the application background of aviation industry parts and avoid the accuracy decline caused by model lightweight,this paper proposes a fine-grained lightweight parts recognition algorithm framework based on algorithm application scenarios.Firstly,based on the improved model,the algorithm designs a fine-grained knowledge distillation framework,and uses the dark knowledge in the teacher network to guide the students to learn object features.Secondly,the isomorphic distillation method is adopted,which can achieve better performance of the student network with a smaller number of parameters while retaining the overall framework of the original model.Finally,the fine-grained knowledge distillation loss is designed,and the hyperparameters most suitable for knowledge distillation of the part dataset are found through the comparative experiment of temperature T and coefficient,and the loss function effectively balances the guiding role of the teacher network and the learning trend of the student network.Compared with other fine-grained algorithms,the improved fine-grained recognition model proposed in this paper improves the learning ability of the model and greatly improves the recognition accuracy of objects.In addition,the knowledge distillation framework based on fine-grained recognition designed in this paper can train a lighter model,which can effectively reduce the model volume while maintaining recognition accuracy.Experiments on public datasets and self-made parts datasets fully prove that the model has good network performance,and the proposed model can be effectively applied to industrial part identification.