Fast and accurate green pepper detection in complex backgrounds via an improved Yolov4-tiny model

In agricultural production, the branches and leaves of green peppers are severely blocked due to the dense plant distribution. This makes the identification of green peppers difficult. Traditional green pepper detection methods entail the problems of low accuracy and poor robustness. This paper introduces a deep learning target detection algorithm based on Yolov4_tiny for green pepper detection. The backbone network in the classic target detection algorithm model is used to ensure classification accuracy. This paper introduces an adaptive spatial feature pyramid method that combines an attention mechanism and the idea of multi-scale prediction to improve the recognition effect of occluded and small-target green peppers. Finally, the method was applied to a test set of 145 images (the target number of green peppers was 602). The AP value of green peppers reached 95.11%; the precision rate was 96.91%, and the recall rate was 93.85%. In order to verify the effectiveness of the module in improving detection performance, we conducted independent combined experiments on the improved module and compared the results with five classic target detection algorithms: SSD, Faster-RCNN, Yolov3, Yolov3_tiny, and Yolov4_tiny. The comparisons verified that the detection rate of the model reached that of the current state-of-the-art technology (SOTA) green pepper detection models. In addition, this green pepper detection model is suitable for real-time detection and embedded development needs of agricultural robots. Such new methods are key components of the technology for predicting green pepper parameters and intelligent picking in unmanned farms.