Deep learning based Customized Model for Features Extraction

摘要

In the last few years, deep learning is one of the powerful technique, used for image recognition. Among the various kinds of deep neural networks, Convolutional Neural Network (CNN) nowadays has been widely used for the purpose of image recognition. In this work, we present the customized model developed for feature extraction of medical packages. Now a days, it is difficult to distinguish an original medicine from counterfeit. The proposed model based on CNN can be useful in identifying the original medicine which forms the first step in the process to identify counterfeits. The medicine images are used as dataset for feature extraction and image classification. Medicine package shape, edges and colour are used for feature extraction of customized model. Classification is done to distinguish medicines which is of same colour and differs in their name, printed text, barcode and the company logo. One of the popular pre-trained CNN architecture model VGG-19 is used for comparing the results of developed customized model. Customized model consists of only five layers convolution layer, maxpooling layer, dropout layer, flatten layer and dense layer. In comparison to pre-trained VGG-19 model customized model reduces number of layers from 19 to 5. Number of layers are reduced to 5 because increasing the number of layers were not showing much improvement in the testing results for given dataset. Training accuracy of 93.17%, validation accuracy of 88.68% and testing accuracy of 76.67% is obtained for the customized model. The results can be made more precise and accurate by optimizing the number of layers in the model. In the proposed model we have used 5 layers. The optimization of number of layers is done based on the prediction accuracy. The accuracy of less than 50% was achieved using 3 layers and it is increased up to 76% with 5 layers. No further improvement in the prediction accuracy was observed by increasing number of layers to 6 or 7, so for the proposed model 5 layers are selected.