Recognition Method of No-seedling Grids of Trays based on Deep Convolutional Neural Network

In the process of replanting, it is necessary to recognize that whether the tray grids are short of seedlings. In the seedling cultivating process, the phenomenon of seedling leaves intruding the adjacent grids due to various reasons make it difficult to identifying the tray grid that if it is lacking seedlings. The traditional image processing methods face the problems of poor human selection features and complex feature extraction process. In this paper, the recognition method of no-seedlings grids of tray based on deep convolutional neural network (GoogLeNet model) is proposed, which uses 9 Inception-V3 modules to construct deep convolutional neural network. Inception-V3 modules use selective multi-scale convolution kernels to extract grid seedlings features, use the convolution kernels factorization to reduce the operation parameters and use Batch Normalization algorithm to improve the model training speed to build the predictive recognition model of tray no-seedlings grids. To verify the effectiveness of this method, this research designed an images acquisition device, collected 3400 pepper tray grids image samples, label the samples by manual. Adjust the model parameters, Build the optimal deep convolution network recognition model when the learning rate is 0.01, Batch size is 100, iteration is 6000, the recognition average accuracy was 94.5%. By using this deep convolutional neural network model, test the samples with and without a leaf or several leaves intrusion individually. The test accuracy of samples without intrusion was 100%. The average accuracy of seedling identification with intrusion conditions was 81.5%. Compared with the traditional image processing method, the average accuracy of recognition with this method was increased by 16%. The results of this research prove that this deep convolutional neural network recognition method is very accurate when there is no intrusion, and the accuracy is relatively high when there is a leaf or several leaves intrusion. This research provide a new theoretical solution for the recognition of no-seedlings grids of tray.