Research on multi-cluster green persimmon detection method based on improved Faster RCNN

Abstract

To address the problem of accurate recognition and localization of multiple clusters of green persimmons with similar color to the background under natural environment, this study proposes a multi-cluster green persimmon identification method based on improved Faster RCNN was proposed by using the self-built green persimmon dataset. The feature extractor DetNet is used as the backbone feature extraction network, and the model detection attention is focused on the target object itself by adding the weighted ECA channel attention mechanism to the three effective feature layers in the backbone, and the detection accuracy of the algorithm is improved. By maximizing the pooling of the lower layer features with the added attention mechanism, the high and low dimensions and magnitudes are made the same. The processed feature layers are combined with multi-scale features using a serial layer-hopping connection structure to enhance the robustness of feature information, effectively copes with the problem of target detection of objects with obscured near scenery in complex environments and accelerates the detection speed through feature complementarity between different feature layers. In this study, the K-means clustering algorithm is used to group and anchor the bounding boxes so that they converge to the actual bounding boxes, The average mean accuracy (mAP) of the improved Faster RCNN model reaches 98.4%, which was 11.8% higher than that of traditional Faster RCNN model, which also increases the accuracy of object detection during regression prediction. and the average detection time of a single image is improved by 0.54s. The algorithm is significantly improved in terms of accuracy and speed, which provides a basis for green fruit growth state monitoring and intelligent yield estimation in real scenarios.