基于优化PSO-BP的多特征融合图像识别算法研究

Research on Image Recognition Algorithm Based on Optimized PSO-BP in Multi-Feature Fusion

为提高水果种类识别的准确性,本文提出一种基于优化粒子群结合BP神经网络的识别算法。在算法初期,针对不同种类水果图像样本,借助K均值聚类分割算法,融合彩色信息和灰度信息,完成目标图像的准确分割,提取目标区域在HSV颜色空间下非均匀量化后的颜色特征,使用分块局部二值模式和灰度共生矩阵,分别提取局部和全局纹理特征,并对与粒子群算法(particle swarm optimization,PSO)结合的BP神经网络进行优化,以获得最优的BP神经网络权值和阈值,同时使用分块的LBP算子和GLCM方法,采用Matlab 2017b软件,对苹果、草莓、柠檬3种水果图像局部和整体纹理信息进行提取,并与传统的PSO-BP神经网络、IPSO-BP神经网络及单一BP神经网络训练之后对测试样本的识别相对误差进行比较。研究结果表明,虽然标准PSO-BP算法对图形的分割效率和识别能力不能与深度学习结果相媲美,但在优化后的PSO-BP中,将3种水果识别率与RCNN系列的优化结果相比并不逊色,且与结合ResNet的SSD算法的结果对比中表现出优异性。该算法保证了图像分割目标的完整性,有效控制了整体算法的时间性,提高了识别过程的精确性。该研究对水果识别精度的提高具有重要的应用价值。

In order to improve the accuracy of fruit type recognition,an identification algorithm based on Particle Swarm Optimization(PSO)combined with BP neural network was proposed.In the early stages of the algorithm,for the different kinds of fruit image samples,the K-means clustering segmentation algorithm is used to fuse the color information and the gray information to complete the accurate segmentation of the target image.Then,the non-uniformly quantized color features of the target area under the HSV color space are extracted.It uses the local binary pattern and gray level co-occurrence matrix to extract local and global texture features respectively,optimizes the BP neural network combined with particle swarm optimization(PSO)to obtain the optimal BP neural network synaptic weights and bias,and uses Matlab 2017b software to extract the local and overall texture information of the three fruit images of apple,strawberry,and lemon with the block LBP operator and GLCM method at the same time.Then it compares the relative errors of the test samples obtained by the traditional PSO-BP neural network,IPSO-BP neural network and single BP neural network.The research shows that the segmentation efficiency and capacity of discernment of the standard PSO-BP algorithm is not as good as the results of deep learning,but the recognition rate of the three fruits by the optimized PSO-BP is superior to the results of the RCNN series and the SSD algorithm combined with ResNet.The algorithm ensures the integrity of the image segmentation target,effectively controls the timeliness of the overall algorithm,and improves the accuracy of the recognition process.This research plays an important role in improving the accuracy of fruit recognition.