Fruit And Vegetable Recognition System Based On Deep Dense Convolutional Neural Network

Object recognition requires the dual task of fusion location and classification,and the implementation process and optimization problems are relatively complex.At the same time,fruit and vegetable images have the characteristics of diverse types,different shapes,and complex detection environment.There are still many aspects of improvement space for fruit and vegetable recognition effect.In this thesis,the deep dense convolution neural network is introduced into the SSD network structure.With the help of the excellent mining ability of the dense network for the effective features and the multi-scale detection structure of SSD network,the detection of multiple fruits and vegetables in the complex background image is well realized.The work done in this thesis is as follows:(1)Design dense weighted connection module.In this thesis,the dense module of dense convolution neural network is improved,and the traditional direct connection mode is changed to the layer by layer connection mode with weight decreasing near the outer layer,which can not only ensure the effective inter-layer connection,but also reduce the accumulation of redundant features to a certain extent.(2)Construct a dual-type module cascade network.In the dense convolution neural network constructed by multiple weighted dense modules,the residual module is introduced.Through multiple development and combination of dense modules,the key and effective features are mined as many as possible.And then the residual module is used to accommodate,digest and eliminate the interference of redundant features,so as to achieve a more stable balance between extraction and elimination of redundancy.(3)Improve the SSD network structure.The feature extraction network of the traditional SSD structure was replaced by a dual-type module cascade network,and the output nodes of the feature map and the size of output features of multi-scale detection were readjusted.This operation not only deepens the number of network layers,but also enhances the connection between the front and back layers,and then improves the accuracy of target detection in recognition network to a certain extent.In this thesis,the network construction is realized on the deep learning Keras library,and the training and testing are carried out on the produced data set with other classic networks.Through the comparison of effects,it can be found that the network designed in this thesis has a more prominent trade-off effect on the recognition accuracy and detection speed.It can realize the detection of 28 frames per second,with an average accuracy of78.45%,which is 4.09% higher than SSD-300.And the system is more robust and more sensitive to the appearance of similar targets.At the same time,the effectiveness of the proposed innovation point is verified by using control variables,which more significantly proves that the improvement of the basic network in this thesis has a better promotion effect.