Affiliations: [a]
Key Laboratory of Grain Information Processing and Control (Henan University of Technology), Ministry of Education, Zhengzhou, PR China | [b]
Henan Key Laboratory of Grain Photoelectric Detection and Control (Henan University ofTechnology), Zhengzhou, PR China | [c]
Anhui Gaozhe Information Technology Co., Ltd
Correspondence:
[*]
Corresponding author. Weidong Chen, Key Laboratory of Grain Information Processing and Control (Henan University of Technology), Ministry of Education, Zhengzhou 450001, PR China. E-mail: [email protected].
Abstract: Unsound wheat kernel recognition is an important part of wheat quality inspection, and it is also a key indicator to measure wheat quality. Research on unsound wheat kernel recognition is of great significance to the correct evaluation of wheat quality. The existing researches on unsound wheat kernel recognition are mainly to directly optimize the classical classification networks, and the recognition effect is often unsatisfactory due to insufficient training data. Aiming at the problem that the recognition rate of unsound wheat kernels is not ideal due to the lack of training data, we propose a Transfer Learning Feature Fusion (TLFF) model. The model uses transfer learning and feature fusion to identify unsound wheat kernels. First, feature extraction is performed by deep Convolutional Neural Networks (CNNs) VGG-16 and VGG-19 pre-trained on the large public dataset ImageNet. Then, the features extracted by the pre-trained neural networks are fused and classified through the flattening layer, fully connected layer, Dropout layer, and Softmax layer. We conduct experiments on single model, two-model fusion, three-model fusion, and four-model fusion, and select the three-model fusion scheme to perform this task. Finally, we vote on the output results of the three best fusion models to further improve the recognition rate. The pre-trained models we use are trained on a large public dataset ImageNet. Since the scale of the dataset is very large, these pre-trained models also have good generalization performance for images other than ImageNet dataset. Therefore, although our dataset is small, we can still achieve good recognition results. Experimental results show that the recognition performance of the TLFF model is significantly better than the existing unsound wheat kernel recognition models.
