Cucumber Flower Detection Based on YOLOv5s-SE7 Within Greenhouse Environments

In the process of cucumber cultivation, the quantity and appearance time of cucumber flowers are important factors influencing the final yields. However, it is labor-intensive to timely record and count the flowers. An optional method is to identify and count cucumber flowers automatically using computer vision technologies based on images taken by cameras installed in greenhouses. However, there are severe problems in images taken with a large field of view in a greenhouse environment, that is, foreground-background imbalance, which renders significant gaps between the detection accuracy and the application requirements even for state-of-the-art computer vision models like Faster-RCNN, SSD and YOLO. This problem can be improved by providing specific datasets and suitable models. Hence, in this paper, two cucumber flower datasets with a wide and medium field of view in greenhouses are constructed. Four attention mechanisms: SE, CA, CBAM and SimAM, are compared and incorporated into YOLOv5s algorithm to improve the detection performance of cucumber flowers during growing states in the greenhouse. The results indicated that our improved model with a SE attention mechanism reached the highest recognition rate than other three methods. The AP@.5 value of the YOLOv5s-SE7 model reached 0.905, which was 3.5% higher than that of the benchmark model YOLOv5s. Meanwhile, it outperformed other state-of-the-art methods such as Faster-RCNN and SSD. The classification detection results of cucumber flowers in three stages, namely bud, bloom and faded flower, reached as high as 0.847 in mAP, suggesting that the proposed model had a good effect in application.