Banana Tree Growth Detection Based On Attention Mechanism And Multi-scale Feature Fusion

Banana is one of the main food crops in China.In recent years,due to serious problems such as natural disasters affecting bananas,it has become particularly urgent to achieve intelligent management of banana industrial parks through banana tree growth detection.The main methods for detecting banana growth currently include manual detection,satellite remote sensing,and machine learning to build mathematical models.All of these methods have problems such as poor accuracy,low efficiency,and insufficient accuracy.With the continuous updating of object detection algorithms in the field of deep learning,research on using object detection models to detect crop growth is constantly emerging.With the main shortcomings being high false detection rate and slow detection speed.After background investigation,there are currently few achievements in the field of banana growth target detection.Therefore,this article aims to lighten the model and reduce the false detection rate.YOLOv5,which has emerged in the past two years and has become stable after multiple versions of updates,is selected as the original model for banana tree detection.Regarding the three classifications of banana tree growth and development stages as the research object,the YOLOv5s model structure is improved,and the SMBi-YOLOv5s model is proposed,which reduces the number of model parameters by 76.6%,The accuracy reached 93.8%,an improvement of2.1% compared to yolov5,and the mAP reached 93.9%.The main research content of this article is as follows:1.Propose the SMBi-YOLOv5s model.Introducing the simAM attention mechanism eliminates the need to introduce additional parameters to the original network compared to traditional attention mechanisms,and instead infers the 3D attention weights of the feature map in one layer;Subtract 20 * 20 scale related feature layers from the YOLOv5s basic model structure to improve model performance;The original C3 module was lightweight processed using the CSPnet channel dimensionality reduction method,which improved the training speed and slightly reduced the average accuracy;Change the CIOU to SIOU to calculate the loss function,further consider the vector angle between the real box and the prediction box,and improve the speed of training and the accuracy of reasoning;Improve the feature fusion network PANET to BiFPN,enhancing the network’s feature fusion ability for targets of different scales.2.Introduce leaf counting and category proportion judgment to assist in correcting growth detection results.In the detection of banana tree growth,due to the blurry boundary between the characteristics of vegetative growth period and flower bud period,the improved model may misdetect the vegetative growth period and flower bud period in the detection.Using the SMBi-YOLOv5s model to detect and count leaves,setting a confidence threshold to determine whether the detection results require secondary judgment,calculating the number of leaves contained in a single tree for classification results below the threshold,and correcting the banana tree classification detection results;Simultaneously add category quantity proportion judgment.3.Design pyqt5 interface based on improved YOLOv5s model.To verify the applicability of YOLOv5s,the pyqt5 framework and Qtdesigner interface design tool were used to develop a graphical interface for banana growth detection,achieving simplified autonomous training and offline image detection functions,and achieving good detection performance.