Greenhouse cultivation can improve crop yield and quality, and it not only solves people's daily needs but also brings considerable gains to the agricultural staff. One of the most widely cultivated greenhouse crops is tomato, mainly because of its high nutritional value and its good taste. However, there are a number of anomalies for the tomato crop that pose a threat for its greenhouse cultivation. Detection of tomato anomalies in the complex natural environment is an important research direction in the field of plant science. Automated identification of tomato anomalies is still a challenging task because of its small size and complex background. To solve the problem of tomato anomaly detection in the complex natural environment, a novel YOLO-Dense was proposed based on a one-stage deep detection YOLO framework. By adding a dense connection module in the network architecture, the network inference speed of the proposed model can be effectively improved. By using the K-means algorithm to cluster the anchor box, nine different sizes of anchor boxes with potential objects to be identified were obtained. The multiscale training strategy was adopted to improve the recognition accuracy of objects at different scales. The experimental results show that the mAP and detection time of a single image of the YOLO-Dense network is 96.41% and 20.28 ms, respectively. Compared with SSD, Faster R-CNN, and the original YOLOv3 network, the YOLO-Dense model achieved the best performance in tomato anomaly detection under a complex natural environment.
Keywords: DenseNet; deep learning; object detection; plant diseases recognition; real-field scenarios.
Copyright © 2021 Wang and Liu.