Identification of Saline Soybean Varieties Based On Trinocular Vision Fusion and Deep Learning

Abstract

Soybean variety recognition is the basis of soybean agronomic yield and commodity attributes. In order to more comprehensively study the recognition performance of deep learning networks under multi-camera fusion, this paper innovatively proposes two new strategies for deep learning of soybean strain recognition based on three-camera fusion. One is image layer fusion and the other is feature layer fusion. Three cameras are used as the experimental trinocular vision. These strategies were evaluated with seven different deep learning network models, including Alexnet, Googlenet, Resnet34, Resnet50, Mobilenet, Shufflenet, and Densenet. Experimental results show that the network performance of both fusion strategies improves with the number of cameras. Notably, Densenet outperforms the other network models. Under the image-layer fusion strategy, Densenet achieves a validation accuracy of 0.9831 and a test accuracy of 0.9938 when three cameras are used. In the feature-layer fusion phase, Densenet achieves a validation accuracy of 0.9875 and a test accuracy when three cameras are used. In the three-camera setup, the image-layer fusion achieved a precision of 0.9729, a recall of 0.9500, and an F1 score of 0.9744. The feature-layer fusion achieved a precision of 0.9756, a recall of 0.9474, and an F1 score of 0.9474. Additionally, based on this research, a new mobile application called “Soybean Seed Classifier” was designed and developed. The results of the study provide a new method for comprehensive soybean seed identification, and the developed software shows practical value in soybean seed identification and breeding processes.