The Improved Mango Plant Detection Model Based on Attention Module Mechanism

Agriculture is one of the sources of income a region can rely on to support its economy. Traditional agriculture relies primarily on human performance and observation, resulting in greater production costs and, subsequently, higher selling prices. Artificial intelligence-based technology can be used to reduce production costs, increase productivity, and provide consumer convenience. An indicator that is easy to interpret in measuring the quality and optimization of plant growth is the visualization of the condition of the leaves. The artificial intelligence technique that can be implemented in this regard is the object detection model. However, the challenge is the complex, multi-object, and multi-intersection condition of the leaves, which causes the model to be less optimal in conducting classification and detection tasks regarding whether the leaf condition is good or not. A YOLOv7 model will be employed in order to detect leaf quality, whether in an “optimal” or “not optimal” condition. To enhance the model's performance by improving accuracy through feature extraction enhancement, YOLOv7 will be integrated with the attention module, called the convolutional block attention module (CBAM). The case study in this research is detecting a mango plant which is one of the plants that can provide a high economic impact and the object observed is the mango plant leaf. Several previous studies related to the implementation of attention modules in object detection include the improved pest-YOLO for real-time pest detection by combining YOLOv3 with efficient channel attention (ECA) and a transformer encoder. The ECA module and transformer encoder were integrated into the backbone and neck block systems of YOLO [1]. The lightweight YOLO model combined with SE-CSPGhostnet by improving the backbone block which employs squeeze-and-excitation networks (SENet) and a convolution technique consisting of regular convolution and ghost convolution [2]. There is a highlighted improvement of YOLOv7 compared to the previous version of YOLO, which is Extended Efficient Layer Aggregation Networks (E-ELAN). YOLOv7's learning ability is enhanced by using this network while maintaining the transition layer's architecture. E-ELAN enhances the structure of computational blocks. E-ELAN can keep its transition layer while changing only the architectural component of the computational block. This allows the model to learn more different features. The fundamental purpose of model scaling is to alter certain model features in order to generate models at various scales to fulfill the needs of different inference speeds.