Identification of green pepper (Zanthoxylum armatum) impurities based on visual attention mechanism fused algorithm

Abstract

Hitherto, assessing the quality of green pepper via identification of impurities has, generally, been done manually. However, manual identification is commonly time and labor intensive. This investigation, thus, taking detection accuracy and reasoning speed on testing dataset as indicators, to explore an appropriate Convolutional Neural Network (CNN) for the detection of green pepper impurities. In terms of detection accuracy, the YOLOv5m outperformed representative target detection algorithms, composed of Faster R-CNN, Grid R-CNN, RetinaNet. Accordingly, the YOLOv5m was further, modified, via the usage of a Similarity-based Attention Mechanism (SimAM) module, to achieve better performance. Fortunately, to compare with YOLOv5m, the average precision (AP) and F1 score for all classes, YOLOv5m-SimAM fused algorithm achieved better results. Furthermore, under the situation of generally same model, parameters, and FLOPs sizes, the inference time of YOLOv5m-SimAM was, unbelievably, 50 % less than that of YOLOv5m. Corporately, both the detection accuracy and reasoning speed of YOLOv5m-SimAM were better than YOLOv5m, especially in reducing inference time. In practice, this case study may mark a critical step forward towards the detection of green pepper impurities to evaluate its quality, from theoretical to application.