Research On Machine Learning-based Methods For Fruit Defect Detection

The presence of defects in the fruit can affect the taste and health of the consumer.So far,the detection of defects in fruit at home and abroad has mainly been done by manual visual discrimination so that defective fruit can be removed.This process is not only time-consuming and inefficient,but the accuracy of the selection of defective fruit is limited by the subjective judgement of farmers and the results may not meet national food safety requirements.Due to the rapid development of computer technology and the qualitative leaps in detection technology that have made farmers’ work easier,technologies such as image recognition are increasingly being used in agriculture,particularly in the identification and detection of fruit defects.Defect detection through machine learning has become an important development trend,and this research is of great significance for the quality of fruit reprocessing in China and for maintaining national food safety.Machine learning theory is used in this paper to detect defects in fruits.Since the methods of defect detection for various fruits are relatively similar,a detailed and systematic description of the detection methods is given in the detection process,taking peanuts as an example,and its main research content is shown below:(1)In order to meet the requirements of peanut fruit detection,a device is designed for automatic defect detection of peanut fruit,an experimental platform for peanut fruit defect detection is established and the design of the main modules is described in detail.Suitable industrial cameras,lens models and other related equipment are identified in the image processing module.At the same time,according to the characteristics of the peanut’s shape and length,a device structure is designed that will allow the peanut to be turned over during transport,which can better obtain all the information on the surface of the peanut fruit.(2)Firstly,the peanut images were data enhanced by spatial geometric transformation and adding noise,and 3600 image datasets of peanut fruits were constructed.According to the characteristics of peanut fruit surface defects,RGB colour space separation of peanut fruits with different defects using colour space analysis was proposed for feature extraction,which accurately achieved defect recognition of diseased and damaged peanuts.(3)This paper proposes a method to optimise the Alex Net network for defect detection in peanut fruit images,which can suffer from slow convergence and low detection and recognition rates.The network structure is optimised by changing the size of the convolutional kernel and adding batch normalisation.Experimental validation is carried out for the Le Net network using small convolutional kernels instead of large convolutional kernels,and reducing the input image size.The results show that the detection of peanut fruit defects can be improved more effectively by improving the method of AlexNet network.