Lightweight target detection for the field flat jujube based on improved YOLOv5

• Published in: Computers and electronics in agriculture 2022-11, Vol.202, p.107391, Article 107391
• Main Authors: Li, Shilin, Zhang, Shujuan, Xue, Jianxin, Sun, Haixia
• Format: Article
• Language: English
• Subjects: Automated picking; Convolutional neural network; Deep learning; Target identifying; YOLOv5
• ISSN: 0168-1699; 1872-7107
• DOI: 10.1016/j.compag.2022.107391

•The flat jujube is a new type of jujube which has four characteristics as flat circle, small target, near-color background and cluster growth.•The method screens out the multi-scale structure of detection network.•Using dual coordinate attention mechanism to improve the ability of feature extraction.•Using bi-directional feature pyramid network to achieve the multi-scale feature fusion.•Depth-separable convolution and ghost model are introduced to achieve the lightweight operation. The efficient detection of the flat jujube in a complex natural environment has great significance in intelligent agricultural operations. Aiming at the problems of the low detection efficiency of field flat jujubes and complex target detection algorithms that are difficult to deploy on low-cost equipment, an improved lightweight algorithm based on You Only Look Once (YOLOv5) is proposed. First, the method screens for the multiscale detection structure that is suitable for the flat jujube by adjusting the number of layers of target detection, which improves the accuracy of detection and reduces the nuisance parameter. Then, multiscale feature fusion is achieved more efficiently by using the bidirectional feature pyramid network (BiFPN), and the feature extraction capability of the model is further improved by introducing a dual coordinate attention mechanism. Finally, the method reduces the difficulties of the model by introducing depthwise separable convolution and adding a ghost module after upsampling layers. The experimental results showed that the mean average precision (mAP) and model size of the lightweight network reached 97.2 % and 7.1 MB. Compared with the YOLOv5 baseline network, the parameters decreased by 49.15 %, while the mAP increased by 1.8 %. The method further improved algorithm performance and reduced computational cost compared with the mainstream one-stage target detection algorithms of the YOLOv5s, YOLOx_s, YOLOv4, YOLOv3 and single shot multibox detector (SSD). Compared to these algorithms, the mAP of the proposed improved model increased by 1.8 %, 0.9 %, 5.5 %, 6.5 % and 2.9 %, respectively. Meanwhile, the model size was compressed by 49.15 %, 73.99 %, 94.42 %, 94.24 % and 86.69 %, respectively. The improved algorithm has higher detection accuracy, while reducing the calculations and parameters, which reduces the dependence on hardware and provides a reference for deploying automated picking of the field flat jujube.