Identifying daylily in complex environment using YOLOv7-MOCA model

Daylily is one of the most favored crops with the great nutritional and economic value. Manual harvesting cannot fully meet the large-scale production at present, due to the inefficient and costly tasks. Mechanized picking can be expected for the large-scale development of daylily industry. However, it is still lacking on the suitable machinery and equipment for the large-scale automated harvesting, due to the complex environments, such as mutual shielding of daylily, interference of weeds, and changes in lighting. It is a high demand for the target identification of daylily under the complex environment for the intelligent harvesting in an automatic harvesting robot. In this study, an improved YOLOv7-MOCA model was established to identify the daylily under the complex environment. The image data of daylily samples were collected from the real harvesting scene in daylily farms. A daylily dataset was established with 12000 sample images, including 2000 raw sample image data of single-target, multi-target and occlusion type in sunny and cloudy weather, and 10000 image data processed by mirroring, brightness transformation, Gaussian filtering, and Gaussian noise processing. Sample images was then enhanced for the better generalization ability of image training models. The daylily dataset in the experiment was divided into a training set, a test set and a validation set, with 7200, 2400, and 2400 images, respectively. The training and identification experiments were then performed on the daylily dataset. The detection effectiveness was compared among the three classical object detection models YOLOv7(You Only Look Once), Faster R-CNN (Fast Region-based Convolutional Network), and SSD (Single Shot MultiBox Detector). The results show that the comprehensive performance of YOLOv7 was better than those of the other two models. As such, an object identification model (YOLOv7-MOCA) was proposed to recognize the daylily in complex environment using YOLOv7. Firstly, a lightweight network model was constructed using MobileOne network as the backbone feature extraction network. Deep convolution was used to replace the ordinary convolution in MobileOne for the less cost of network computing that caused by feature extraction and object detection. The feature extraction efficiency and model detection speed of the model backbone network were compared with the replacement of two mainstream lightweight neural networks. Secondly, the Coordinate Attention (CoordAtt) attention mechanism was integrated in the neck network, in order to improve the receptive field of the model before the convolution operation. The mechanism was enhanced the attention of the model in the characteristics of small target samples of daylily. The comprehensive performance of intelligent identification was achieved in the real-time detection effectiveness of daylily samples. Finally, a YOLOv7-MOCA lightweight daylily fast identification model was designed by ablation test. The MobileOne was used as the backbone network, whereas, a neck network was utilized the attention mechanism of CoordAtt. The experimental results showed that the model YOLOv7-MOCA detection accuracy reached 96.1% with the recall rate of 96.6% and F₁ coefficient of 0.96. The model weight was 10 MB, whereas, 58 frames per second was used to transmit. Compared with the YOLOv7, the detection accuracy of the improved model increased by 3.2 percentage points, the average detection speed increased by 26.1%, whereas the weight was reduced by 86.7%, indicating the small model weight and fast identification speed. The optimal parameters were obtained from the improved lightweight network model. The detection speed and model weight were significantly redesigned to effectively optimize the identification and detection of daylily in the complex environment of the network model, in terms of the mutual occlusion of target samples, weed interference, and light change. This improved model can be expected to intelligently identify some targets under complex environments for the automatic harvesting equipment of daylily. The finding can provide the technical support for the research of intelligent harvesting equipment of daylily.