| Summary: | Impurities in wheat seriously affect wheat quality and food security. They are mainly produced during the operational process of combine harvesters. To solve the recognition and classification problems associated with impurities in wheat, a recognition method using an improved convolutional neural network is proposed in this article. A labeled dataset of normal wheat and five impurities is constructed, using which the Wiener filtering algorithm and the multi-scale Retinex enhancement algorithm are employed for image preprocessing. Based on network research using Inception_v3, improvement and optimization are undertaken before designing the WheNet convolutional neural network, which is intended for automatic recognition of wheat images. Under the same conditions, comparative experiments using the WheNet, ResNet_101 and Inception_v3 networks are conducted. Indexes such as receiver operating characteristic, area under curve (AUC), and recall rate are adopted to evaluate the experimental outcomes. Experimental results indicate that the WheNet network achieved the most efficient results. It also shows a shorter training time, and its recognition accuracies for Top_1 and Top_5 of the test set are 98.59% and 99.98%, respectively. The mean values of both the AUC and recall rate of the network on the recognition of various images of impurities are higher than those of the ResNet_101 and Inception_v3 networks. Consequently, the WheNet network can be a useful tool in recognizing impurities in wheat. Furthermore, this method can be used to detect impurities in other fields.
|
|---|
