Research On Deep Learning-Based Jimu Part Classification

As a comprehensive technology of machinery,electronics and computer vision,industrial automation technology has been attracting extensive attention and developed rapidly recently years.Image classification is a challenging task in the field of industrial automation and has great application potential.The traditional image classification method uses various feature extraction algorithms to extract features,and analyzes the feature similarity between different images to determine the labels.With the rapid developing of image processing and deep learning,many image classification algorithms based on neural networks have been proposed.The effects of these algorithms have been fully verified in some large open data sets,but it is still difficult to evaluate their effects in the actual industrial environment.Due to the complexity of industrial environment conditions,it is difficult for existing algorithms to work effectively in practical applications.More importantly,there is a lack of open industrial environment data set to evaluate these algorithms.Therefore,this paper establishes a dataset based on the real industrial environment for the classification of building blocks encountered in the actual industrial production.The data set established in this paper is used for the qualitative and quantitative comparison of four classical neural network algorithms.Finally,a building blocks classification network model based on the multi-model voting mechanism is proposed.The main research contents of this paper can be summarized as follows:(1)We propose a benchmark dataset in this paper,named FIST-Toy,which are collected in real industrial environment.The FIST-Toy can be applied to many problems,such as target classification,recognition and location.The FIST-Toy contains 2297 images of 10 different categories,including different color of the same contour blocks parts,the same color contour blocks parts and components,and same color contour similar symmetrical parts.All the images are collected in different time,so as to the color of them are difference,the color changes and symmetric block parts can better evaluate the performance of image classification algorithms.(2)This paper has conducted a large number of experiments on several mainstream CNN-based models,and designed a basic classification framework for the classification of different building block toy parts.Under the same experimental conditions,the performance of AlexNet,VggNet-16,Inception-v3 and ResNet-34 is analyzed.These mainstream CNN models cannot effectively solve the problem of the classification of symmetrical building blocks,which also provides some new perspectives for the research of the classification task.(3)Through the analysis of the experimental results,it is found that the ResNet-34 network model is superior to other network models in the classification task of building block toy parts.In this paper,a multi-model voting mechanism-based block part classification model is proposed,and the experimental results show that the classification effect of this model is superior than the existing CNN network model.