Fruit image recognition and classification method based on improved single shot multi-box detector

In recent years, pattern recognition has been gradually applied to the field of agriculture, especially fruit classification and rating based on image. However, the target fruit is affected by the interference factors such as light change, uneven brightness, similar background, branches and leaves, shadow coverage, etc. In order to solve the problems of low recognition rate and low generalization of fruits in the natural environment, we proposed an improved SSD (Single Shot Multi-box Detector) to classify apple, persimmon, nectarine and pear. The original network is optimized and the more advanced soft NMS (Non-Maximum Suppression) is used to obtain the anchor boxes, so as to provide a better initial value for the prediction of the bounding box. At the same time, under the premise of keeping good performance independent of the network architecture, batch normalization is used to initialize the de-random training model, which brings stable and predictable gradient of the detector. Through the identification test of four kinds of fruits collected in the natural background, the experiment shows that the improved SSD fruit detection model is effective. The fruit recognition speed is the fastest and the detection accuracy is the highest compared with other methods. The average detection accuracy in various environments has reached 92.4%. It can be concluded that the proposed method will realize the accurate detection of many kinds of fruits, and provide a new scheme for the fruit recognition and detection problems in agricultural automatic picking.