颗粒状农作物彩色色选图像处理算法研究
哈尔滨工业大学工程硕士学位论文
Abstract
Color selection equipment is an application of machine vision in the field of industrial and agricultural sorting. At present, the color sorting system usually adopts a DSP+FPGA embedded hardware platform, and high-resolution images are used to classify by setting a threshold. The platform needs enough storage space, and there are disadvantages that the classification accuracy depends on experience. In recent years, the development of GPUs has provided the hardware basis for the application of complex image processing algorithms in color-selection systems. This paper combines the DSP+FPGA platform and the CPU+GPU platform to study the image compression algorithm, image classification algorithm and feature parameter detection algorithm selected by color.
Firstly, a color-selection image processing hardware platform is set up to complete image acquisition, display, and processing. In the DSP+FPGA platform, a high-resolution color linear array CCD camera is used to acquire images. In the FPGA, a dual-port RAM is built to realize the data storage of the Camera Link interface. The VGA is used to control the image display, and the data is transmitted with the DSP. In the CPU+GPU platform, the low-resolution industrial camera and the image capture card are used to complete image acquisition and a TensorFlow deep learning framework is built.
Secondly, the image compression and reduction algorithm in color sorting are investigated. By selecting the wavelet function through comparison and analysis, the wavelet transform image decomposition and reconstruction are realized. Simulation analysis of EZW and SPIHT encoding methods, the results show that the large amount of computation, it is difficult to ensure real-time. This paper combines the image features of granular crops with color sorting, and proposes a simplified wavelet transform image compression algorithm. It directly performs Huffman coding on the low frequency coefficients after wavelet decomposition. The algorithm is transplanted to DSP for testing. The results show that the algorithm has higher compression ratio and real-time performance under the condition of satisfying the image quality.
The color classification image classification algorithm based on deep learning is studied. Peanut image data sets are first established, which are divided into training set and test set. Convolutional neural network algorithm in deep learning is used to implement image classification. In order to improve the classification accuracy and real-time performance, this paper optimizes the data set according to
哈尔滨工业大学工程硕士学位论文
the classification results; it optimizes the convolutional neural network from several aspects such as reducing over-fitting, accelerating the training convergence speed, and simplifying the network structure. The test results show that the optimization program has achieved the expected goal.
Finally, the feature parameter detection algorithm is studied. For low-resolution sample images, two super-resolution reconstruction methods based on deep learning are researched, including SRCNN and ESPCN. According to image reconstruction quality and running time, ESPCN is selected for super-resolution reconstruction. The peanut image is then subjected to color processing, filtering and morphological operations, followed by edge detection and damage detection. Finally, feature parameter detection algorithm is written in C language and Python language, respectively, and transplanted to DSP + FPGA platform and CPU + GPU platform test. The results show that the algorithm can accurately complete the detection of characteristic parameters of peanut edges and damaged areas in real time.
Keywords: image compression, image classification, deep learning, color sorting, super-resolution reconstruction
哈尔滨工业大学工程硕士学位论文
目录
摘要........................................................................................................................ I Abstract ..................................................................................................................... I I 第1章绪论 (1)
1.1 课题研究的目的和意义 (1)
1.2 国内外研究现状 (2)
1.2.1 色选系统研究现状 (2)
1.2.2 图像压缩算法研究现状 (3)
1.2.3 深度学习在图像处理领域的研究现状 (4)
1.2.4 色选系统中特征参数检测算法研究现状 (5)
1.3 本文主要研究内容 (6)
第2章彩色色选图像处理平台的设计 (8)
2.1 引言 (8)
2.2 彩色色选系统整体方案设计 (8)
2.2.1 需求分析 (8)
2.2.2 DSP+FPGA平台总体硬件方案 (8)
2.2.3 CPU+GPU平台总体硬件方案 (9)
2.3 DSP+FPGA硬件平台的搭建 (10)
2.3.1 图像采集 (10)
2.3.2 图像显示 (14)
2.3.3 图像处理 (18)
2.4 CPU+GPU硬件平台的搭建 (18)
2.5 本章小结 (20)
第3章彩色色选图像压缩还原算法 (21)
3.1 引言 (21)
3.2 图像压缩质量评价 (21)
3.3 图像小波变换的分解与重构 (22)
3.3.1 小波分解 (22)
3.3.2 小波分解后子图像灰度直方图 (24)
3.3.3 小波分解后子图像灰度数值分布 (25)
3.3.4 小波函数选取 (26)
哈尔滨工业大学工程硕士学位论文
3.4 基于小波变换的图像压缩编码算法 (28)
3.4.1 嵌入式零数小波编码 (28)
3.4.2 多级树集合分裂编码 (29)
3.4.3 仿真分析 (30)
3.5 小波变换图像压缩还原算法简化与测试 (31)
3.5.1 小波变换图像压缩还原算法简化 (31)
3.5.2 小波变换图像压缩还原算法测试 (32)
3.6 本章小结 (33)
第4章基于深度学习的彩色色选图像分类算法 (35)
4.1 引言 (35)
4.2 卷积神经网络基础理论 (35)
4.3 基于CNN的花生分类算法 (38)
4.3.1 数据集 (38)
4.3.2 卷积神经网络的构建 (39)
4.4 基于CNN的花生分类算法的优化 (41)
4.4.1 数据集优化 (41)
4.4.2 正则化减小过拟合 (43)
4.4.3 指数衰减法设置学习率 (44)
4.4.4 滑动平均模型加快训练收敛速度 (45)
4.4.5 神经网络结构的简化 (46)
4.4.6 综合优化方案 (47)
4.5 基于CNN的花生分类算法的优势 (48)
4.6 本章小结 (49)
第5章彩色色选的特征参数检测算法 (50)
5.1 引言 (50)
5.2 基于深度学习的超分辨重建 (50)
5.2.1 基于卷积神经网络的超分辨技术 (50)
5.2.2 基于亚像素卷积神经网络的超分辨技术 (54)
5.3 特征参数检测算法 (56)
5.3.1 颜色模型 (56)
5.3.2 图像滤波 (57)
5.3.3 形态学运算 (58)
5.3.4 边缘检测和破损区域检测 (59)
哈尔滨工业大学工程硕士学位论文
5.4 特征参数检测算法测试 (60)
5.4.1 DSP+FPGA平台测试 (60)
5.4.2 CPU+GPU平台测试 (61)
5.5 本章小结 (63)
结论 (64)
参考文献 (65)
攻读学位期间发表的学术论文及其他成果 (69)
哈尔滨工业大学学位论文原创性声明和使用权限 (70)
致谢 (71)