Real-time Citrus Detection Based On Improved SSD Deep Learning Model

Accurate and fast fruit detection method is an important prerequisite for intelligent and efficient operation of citrus picking robots,while traditional detection method has limited accuracy and weak generalization ability,especially slow identification speed,which is difficult to meet the requirements of real-time detection of citrus.This paper proposes a real-time detection method for citrus based on improved SSD deep learning model.The main contents are as follows:Firstly,the SSD model was improved and the SSD_MobileNetV2 model was designed.At the basic convolution layer,the lightweight MobileNetV2 network is used to replace the heavily weighted vgg-16 network;at the auxiliary convolution layer,the reverse residual structure is used to replace the traditional convolution structure to serve as the basic structure of the layer;width to height ratio of the prior more in line with the geometric characteristics of citrus fruit is set.Secondly,the citrus image and video data sets were constructed.The image of citrus fruit with different pixel area,different illumination condition and different fruit state is collected.An adaptive histogram equalization method is used to preprocess the bad images.Mirror,rotation and other methods are used for data amplification.In order to improve the ability of the model to detect occluded or overlapping fruits,the foreground regions sample of occluded or overlapping fruits are manually labeled.Video collection is conducted from three angles,namely,elevation,horizontal and overlook,to enhance the diversity of citrus samples.Finally,the citrus detection model was trained and tested.Use the transfer learning method to speed up the training.The total loss function and training parameters are determined,the learning rate attenuation method,momentum term and other methods are adopted to optimize the training process,and the model with the best performance is selected as the citrus detection model in this paper,and the results are compared with the model before the improvement,without the labeling method of the citrus foreground regions and other detection methods.Finally,the experiment of video streaming citrus detection based on raspberry PI device is carried out.The experiment shows that the precision rate,recall rate and average precision of the improved detection model reach 94.68%,88.31% and 85.50% respectively,which are 1.07%,2.14% and 5.08% higher than before.In particular,the detection speed reaches 41.31 FPS,the number of parameters is reduced to 1/6 of the original,and the detection speed is increased by 2.52 times.In the detection of occluded or overlapping fruits,the average precision of the detection model was improved by 3.76% by using the method of labeling the target foreground region.Compared with Faster r-CNN and the method of HOG combined with SVM,the average precision of the detection model adopted in this paper is 7.54% and 17.38% higher and the detection speed is increased by 9 times and 15 times respectively.The transplanting effect of the detection model on the raspberry PI embedded device is favorable,as video stream citrus detection can be basically achieved.The real-time detection method of citrus based on improved SSD deep learning model can improve the detection speed more obviously on the premise of ensuring high recognition accuracy and robustness,which meets the real-time requirements,provides a technical reference for fruit detection method,and has certain application value for improving the performance of picking robot.