An improved YOLOv5s method based bruises detection on apples using cold excitation thermal images

Abstract

Bruising is one of the key factors that causes postharvest losses, which decreases the economic efficiency of fruit. Nevertheless, the detection of bruises still relies mainly on manual work, which is strongly subjective with long labor time and low efficiency. Accordingly, it is necessary to design an efficient fruit bruise detection approach. Thermal imaging (TI) is a fast and effective nondestructive testing technology. However, the commonly applied thermal excitation TI-based bruise detection may lead to a decrease in the shelf life of the fruit. Therefore, this study uses apple as the research object, introduces cold excitation to improve the sensitivity of bruise detection, and then constructs a simple longwavelength infrared range (7.5–13 µm) TI system to acquire the thermal image of bruised apples. In addition, the low signal-to-noise ratio of thermal images also leads to detection performance degradation. Thus, the YOLOv5s network is applied and improved to achieve better detection. The specific methods are described as follows: (1) Since the thermal images have the problem of duplicated RGB data, group convolution is used to reduce the feature duplication computation. (2) The bottleneck structure of YOLOv5s is replaced by the ghost bottleneck (GB), and the number of bottlenecks is reduced to decrease the computational quantity of extracting redundant features of thermal images. (3) The shrinkage module is inserted into the GB, and the threshold is automatically obtained through two fully connected layers without relevant professional knowledge to eliminate noise in the features that may cause performance degradation. The F2 score, mAP and mAP50 of the proposed model are 97.76%, 86.24% and 98.08%, respectively, which are better than those of YOLOv5s. Moreover, the computation and the FPS of the proposed model are 1.31 GFLOPs and 160, which are 31.95% and 121.21% of those of the YOLOv5s, respectively.