Research On Fruit Ripeness Recognition Algorithm Based On Deep Learning

With the development of agricultural intelligence,the identification and picking of ripe fruits are gradually replaced by manual operations by intelligent machines.The core technology of picking robot is also the main research direction of this paper.This kind of robot for picking ripe fruit is to solve the cumbersome,time-consuming and expensive problems in the process of manual fruit picking,and the artificial picking error is large,and it can also effectively Avoid errors.However,the detection of fruits and vegetables by computer vision systems using artificial intelligence is also affected by the background environment.For example,occlusion by leaves and branches,uneven lighting,and other unpredictable factors.Aiming at the above problems,this paper studies the recognition algorithm of fruit ripeness.Taking the images of red apples with different ripeness as the research object,the research and optimization and improvement of the red apple ripeness recognition algorithm are carried out.The research content of this paper includes the following parts:This paper selects Darknet-53,the backbone network structure of the YOLOv3 algorithm,because the normalization of Darknet-53 is not applicable to the small batch characteristics of the apple maturity feature map in this paper,and the normalization method is replaced,and from the perspective of feature reuse,introduce a new Densenet network to enhance the efficiency,at the same time,introduce an attention mechanism before each dense block of the Densenet network,reduce the model’s attention to the interference information such as leaves,and improve the recognition accuracy of the model.The training of the improved model,the preprocessing of the image,the comparative analysis of the experiment,and comprehensively verify the rationality and effectiveness of the improved method in this paper.This paper mainly aims to improve the detection accuracy performance of the fruit ripeness recognition algorithm.Using the deep learning framework,an improved network model named Darknet-53-GN-Densenet is proposed,and an attention mechanism is introduced to use this model.To carry out the identification and detection of apple maturity.First,analyze the Darknet-53 network structure used by the original classic YOLOv3 algorithm.The Darknet-53 network structure has certain drawbacks in the normalization method,that is,it is more friendly to the detection of large-scale feature maps,while the fruit of this paper is The maturity target detection graph is small in scale,and the test set is also a small batch mode.Therefore,the group normalization method is used to replace the original batch normalization,which solves the problem of batch dependence during model training and avoids model training from being affected.The constraints of hardware conditions meet the requirement that the accuracy of the model is not affected during small batch detection.In order to further improve the detection accuracy and real-time performance,on the basis of the improved Darknet-53-GN structure,this paper continues to introduce the Densenet network into the algorithm,and at the same time adds the attention mechanism module.The feature samples have been compared in the training set,validation set and test set.It has been verified that using the improved Darknet-53-GN structure alone,the accurate recognition rate of the test set has been improved to a certain extent.When introducing a new Densenet network,add attention After the force mechanism,the transfer efficiency of model features and gradients has been improved accordingly,the network is more convenient to train,and the recognition accuracy of the algorithm has been greatly improved.It can be concluded that the combined improvement of the three optimization methods improves the detection accuracy performance of the YOLOv3 algorithm in apple maturity target recognition.