Research On Image Classification Algorithm Of Granular Crops Based On Embedded Platform

In the agricultural production process,granular crops such as peanuts,soybeans and corn are prone to skin damage and affect the quality and commercial value of crops.Traditional manual sorting methods require a large amount of labor and are inefficient.In this paper,peanuts were used as representative of granular crops to study the image classification algorithm of granular crops based on embedded platform.Through the software and hardware collaborative design,the image classification algorithm is transplanted and implemented on the Zynq SoC platform.First,design the system platform.The hardware platform is built,including the model training platform with GPU platform as the core and the model inference platform with Zynq SoC platform as the core.Complete the peanut image acquisition and preprocessing,reduce the impact of shadows of crop image caused by uneven illumination of light source,and construct the image data set.Secondly,study the image classification algorithm based on feature extraction and support vector machine(SVM).Aiming at the problem of SVM algorithm lacking the ability to extract image features,the three image feature extraction algorithms of histogram of oriented gradient(HOG),local binary pattern(LBP)and scale-invariant feature transform(SIFT)are compared and analyzed.The HOG algorithm with high accuracy and short running time is used as a feature extraction algorithm for granular crops.According to the characteristics of the embedded platform,the algorithm is simplified and the computational resource overhead is reduced.Then,study an improved convolutional neural network(CNN)image classification algorithm was.Based on the lightweight network SqueezeNet,the network topology structure is improved,and the comprehensive optimization strategies such as Msra parameter initialization,adaptive learning rate attenuation and momentum smoothing are used to optimize the network model training phase,improve the generalization ability of the network model,and accelerate the convergence speed.Finally,propose a reusable FPGA hardware acceleration design method,combining with the principle of algorithm and the characteristics of embedded platform.Use high-level synthesis tools to design accelerated optimization strategies and improve reuse.The convolutional layer,the pooled layer,the fully connected layer,and the radial basis function layer are designed separately in a modular manner.Then set up the experimental environment and test the acceleration effect.The experimental results show that the FPGA hardware acceleration effect is excellent,and the SVM algorithm has the characteristics of good real-time performance and can achieve high accuracy.The CNN algorithm has a better classification effect in the task of granular crops image classification based on embedded platform.