MBNet: A multi-branch network for detecting the appearance of Korla pears

Accurately assessing the appearance properties of pears is crucial to postharvest quality assessment. Conven-tional machine-learning detection methods only extract single features, and most cannot make multiple as-sessments in one shot. In this study, a system using three cameras is leveraged to feed sensory data to a single -shot multi-branch network (MBNet) to detect multiple quality features of Korla pears. The basic MBNet extracts convolutional characteristics, and an appending network that includes four branches distinguishes the four quality features. An integrated loss function is applied for model training, and a strategic transference re -judgement module is applied to ensure the reliability of the final results. The MBNet model has high accuracy and low complexity, and its errors are 5.1, 3.1, 6.4 and 4.8% for transfer status, calyx, protuberance and blush features, respectively. The accuracy of the full system is 89.3% when detecting multiple Korla pear quality features in a single shot, which reduces the model size by 75% and is beneficial to postharvest quality assurance.