A novel and high precision tomato maturity recognition algorithm based on multi-level deep residual network

• Published in: Multimedia tools and applications Vol. 79; no. 13-14; pp. 9403 - 9417
• Main Authors: Liu, Jun, Pi, Jie, Xia, Liru
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.04.2020; Springer Nature B.V
• Subjects: Accuracy; Algorithms; Artificial neural networks; Classification; Complexity; Computer Communication Networks; Computer Science; Convolution; Data Structures and Information Theory; Image detection; Image enhancement; Multimedia Information Systems; Neural networks; Object recognition; Picking; Sensors; Special Purpose and Application-Based Systems; Deep learning; DenseNet; Loss function; Structuring sparse; Tomato detection; Network tailoring
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-019-7648-7

Since the existing tomato picking system uses multispectral sensors, color and other passive sensors for tomato detection and recognition, its detection range is very small, anti-interference ability is also weak, and tomato maturity detection cannot be performed accurately in real-time. How to detect tomato information from the massive image data obtained from tomato picking equipment and improve the recognition accuracy is a challenging research topic at home and abroad. This paper proposes an improved DenseNet deep neural network architecture, and uses it to solve the detection problems of maturity tomato in complex images. In order to enhance the accuracy of feature propagation and reduce the amount of stored data, a structured sparse operation is proposed. By dividing the network convolution kernel into multiple groups, the unimportant parameter connections in each group are gradually reduced during the network training process. In addition, since the dataset constructed in the field of tomato picking has imbalance, we introduce the Focal loss function to identify the tomato in the classification layer so as to enhance the accuracy of the final classification prediction of the tomato detection system. A large number of qualitative and quantitative experiments show that our improved network in this paper is superior to other existing deep models in terms of detection rate and FPPI, and its computational complexity is lower than that of DenseNet algorithm 18% under the same hardware and software configuration.