针对不同型号车辆外观差异较小,车辆检索困难的问题,构建一种两阶段细粒度车辆检索算法.该算法选择包含有效信息的特征,第一阶段通过广义平均池化(Generalized Mean Pooling)产生全局特征描述子,最后通过欧氏距离法得到初次检索结果.第二阶段通过Faster R-CNN预测目标区域的类别得分和位置坐标,在初次检索结果中找到与该查询类别相同的目标区域,并结合扩展查询(Query Expansion)对目标区域特征再次进行欧氏距离计算,检索出最终相似的图像.实验结果证明,该方法在细粒度车型数据集上取得了较好的效果.

针对车辆型号繁多、部分型号间差异较小带来车辆分类困难的问题,构建一种基于改进的Mask R-CNN细粒度车辆型号识别算法.改进后的算法采用聚合残差-特征金字塔网络(ResNeXt-FPN)提取特征图;调整了区域建议网络(RPN)中锚(Anchor)的尺寸大小;用Soft-NMS代替了非极大值抑制算法(NMS),以提高检测精度;去除掩码分支,节省了预测时间.为了验证算法改进的效果,将其与最新的目标检测算法进行对比.实验结果证明,改进的算法提高了车辆识别的准确率,比原始算法准确率提升了2％.

在大米品质检测中,大米图像的分割是其中的必要环节.常用otsu阈值分割算法实现大米图像分割的过程.但是传统一维otsu算法只是基于灰度值信息的分割效果不理想,而传统二维otsu算法的存在计算冗余,过程复杂、分割需要的时间长,为此提出了一种大米图像分割的pso优化二维otsu算法,通过粒子群优化算法实时更新粒子速度和位置,利用其寻优特性能够快速求解得到分割阈值,实现大米图像分割,减少otsu算法自适应阈值的冗余计算,提高运行速度.结果表明该分割算法在大米图像分割实验中,有效降低了二维otsu运算过程的复杂度,减少了图像阈值分割的时间,并且没有对取值结果造成影响,寻找的阈值准确,提高了阈值分割的效率,具有一定的可行性,为大米图像处理研究和粮食品质检测提供了参考.

针对粘连大米图像分割中轮廓的凹点检测问题,提出了一种基于脉冲耦合神经网络(PCNN)和harris角点检测结合的方法.该文简要介绍了基于脉冲耦合神经网络的图像分割、harris角点检测算法.首先对采集的图像进行PCNN图像分割将目标和背景分割;然后提取分割后的粘连大米轮廓,并进行高斯平滑;再根据平滑后的轮廓曲线进行harris角点检测,识别准确的凹点.

wheat hardness is an important index for wheat quality evaluation .In this paper , the wheat grain collision ob-jects produce audio test signal analysis , combined with wavelet transform and discrete cosine transform algorithm for wheat audio test signal processing , using the multiresolution feature of wavelet transform and discrete cosine transform energy compression and related ability , extract useful characteristic parameters of linear regression analysis method is adopted to establish the regression model based on these feature parameters .Experiments show that the built model of the linear cor-relation coefficient can reach 0 .957 3 .

In view of the problem,that the insect damaged wheat kernels' detection worked intensity,lowly and low accuracy,the paper introduced a kind of method about insect damaged wheat kernels based on the basic principle of acoustics.First,the system collected the ultrasonic signals of wheat kernels which hit target of metal by using self-made voice signal acquisition device,then processed the pretreated signals with Fourier transform and extracted the related characteristics.According to the features,the injured wheat kernels were classified through the BP neural network.Through test insect damage and perfect wheat kernels 30 sample respectively,the results show that the correct detection rate of insect damage wheat kernels by using acoustic characteristics was 90%,perfect wheat grain recognition rate was 93%,the overall accuracy of wheat discrimination at 91.5%.

应用小麦粉加工精度测定仪测定了不同面筋含量、不同加工精度的212份小麦粉样品的粉色、麩星含量、最大麩星面积、麩星数量密度、白度,同时测定了小麦粉的灰分、面糊色度.首次发现麩星数量密度参数与小麦的品种无关,与小麦粉面糊的色度线性相关,相关系数0.9以上,预示有可能建立直接由小麦粉评价小麦粉最终产品颜色的方法.

A new acoustic method based on wavelet packet transform was proposed for the wheat hardness detection.The system collected the ultrasound signal of the wheat and exploded it by using the wavelet packet transform,got the sub-band energy of signal and extracted some characteristic parameters.Through linear regression analysis,the regression model between wheat hardness and the characteristic parameters was established.The linear correlation coefficient r2has reached 0.9589.

By processing the ultrasonic signals of wheat with discrete cosine transform,we extract the characteristic parameters what can reflect the moisture content of wheat.Using a linear regression analysis,we establish based on these characteristic parameters of wheat moisture content prediction m0del.The experiments show that the maximum average absolute of the predictive model what is establishment of the same varieties of wheat moisture content is 0.02,the maximum average relative error is-0.018%.

In the unilateral light conditions,due to the effect of the interaction between rice and chalkiness,the adhesion rice image collected will lost the information of the low gray after Binarization easily.Therefore,there is need to enhance the image.This article talks about the enhancement of the touching rice image using transformation of powers,histogram equalization and local enhancement,compare and analyze them,in order to get the most effective method.The results shows that the effect of transformation of powers and local enhancement works best,and conducive to subsequent adhesion split.

基于图像处理技术的小麦粉加工精度测定仪能够快速、客观、直接测定粉色、麸星,测定结果与感官检验一致,能够准确反映小麦粉的加工精度。在小麦粉加工精度实物标样的研制中,多年的应用表明该仪器是指导小麦粉生产,研究小麦粉品质的一种科学、方便的检测工具。

moth-eaten factor is the one seriously affects the quality of rice.Currently,it generally uses vision to detect moth-eaten rice,but human factors affect the detection results.This paper presents a shape-based feature detection method.extract the edge point coordinates by the direction of freeman chain code from the moth-eaten rice image,and then calculate the angle of two adjacent points,analysis the data obtained,and then calculate their mean,variance,third-order moments,entropy and other features,finally,using fisher classifier to classify the moth-eaten rice.Experiments show the rate of the moth-eaten rice is 95.65%.

The ultrasonic signal of the wheat grain collides with metal objects contains a wealth of quality information.So,we extract the characteristic parameters what can reflect the wheat moisture by processing the ultrasonic signal with wavelet transform.With binary linear regression analysis method,we analyze these characteristic parameters,and establish based on these constants of wheat moisture content prediction model.The experiments show that the average absolute error of prediction model is 0.002;the average relative error is 0.012%.

The quality of rice was influenced by injured kernel seriously,rice injured kernel detectionwas very significative.Detecting the rice injured kernel by the method of graphical analysis is a new research subject while rice injured kernel edge detection is one important task.This article use the LoG operator method,frequency domain detection method,morphology to detect worm-eaten rice edge,and compare their detection effect and speed,in order to get the most effective edge detection method.The result shows that morphological detection method was the best,and the processing speed was fastest.

The audio signals of wheat grain colliding with objects contain large amounts of information.By processing the audio signals with wavelet transform technology,some useful feature parameters can be extracted.With linear regression technology,we analyze these feature's parameters,establish relations between these constants and wheat hardness,and build the regression model.The linear correlation coefficient of the regression model has reached0.9505.

针对目前数据库系统多不提供元数据获取及输出的问题,提出一种元数据获取、存储及打印的方法并给予设计和实现,该方法具有一定的实用价值。

The application of acoustic mensuration in the determination of wheat quality is studied.The audio signal of wheat which is caught by the home-made devices is pre-processed by the use of voice signal processing technology.Several commonly used endpoint detection algorithms are compared through simulating experiment of MATLAB.The start point and end point of the audio signals of wheat are detected with the method that the character of LPCC and variance of frequency band are combined,which detected the start and end point of the audio signal of different wheat comparatively accurate and provided the preparation work to the acoustic measurement and the feature extraction of wheat.

The morphological watershed algorithm is applied.The rice images are treated by top-hat transformation, the adhesive degree of touching rice is reduced by morphologic eroding operation,and the adhesive rice is separated by the distance transformation combined with watershed algorithm.The results show that the adhesive rice images can be segmented effectively by morphological watershed algorithm.

To guarantee the security of grain,we should pay more attention to the improvement of the ability of grain production.A fast and accurate measure detecting the appearence quality should also be provided for the purchase and market circulation of the commodity grain.Aiming at the over segmentation and under-segmentation question which easily happened on the segmentation of adhesion food particles image,this paper proposes an improved method of mathematical morphology,which is on the basis of the grain image segmentation designed by Lingyun.Through contrast experiments between the original algorithm and the improved algorithm,the results showed that the improved algorithm can effectively solve the segmentation problem of adhesion particles,specifically the segmentation problem of the smaller touching area of particles and the higher broken rice ratio.Mic-monitor to the drying process can be achieved.

The system collected the sound signals when larvae of grain insects ate grains in a grain bulk,and then processed the signals by a signal processing technique to extract the useful signal features.According to the features,the insects were classified through the artificial neural network.The system collected the sound signals of three kinds of insects,and the classification correct rate was 0.958 3.