摘要：【中文】本申请的目的是提供一种用于车辆外观分割的神经网络模型构建方法与系统。与现有技术相比，本申请通过一种用于车辆外观分割的神经网络模型构建方法，首先获取已对车辆外观标注过的样本图像集，接着将所述样本图像输入改造后的神经网络进行训练获得自适应边距损失函数输出的损失值，其中，所述改造后的神经网络结合了自适应边距损失函数，最后判断所述损失值是否满足预设的阈值，如满足则完成用于车辆外观分割的神经网络模型构建。通过该方法使得车辆外观图像中类别样本较少或更难区分的类别可得到更高权重，很好解决了类别样本不均衡问题，能更准确完成车辆外观的图像分割，该方法在低显存占用情况下对车辆外观分割获得更高的准确度。 【EN】The application aims to provide a neural network model construction method and system for vehicle appearance segmentation. Compared with the prior art, the method for building the neural network model for vehicle appearance segmentation comprises the steps of firstly obtaining a sample image set marked on the vehicle appearance, then inputting the sample image into a modified neural network for training to obtain a loss value output by a self-adaptive margin loss function, wherein the modified neural network is combined with the self-adaptive margin loss function, and finally judging whether the loss value meets a preset threshold value, and if so, completing the building of the neural network model for vehicle appearance segmentation. By the method, the categories with fewer category samples or more difficult to distinguish in the vehicle appearance image can obtain higher weight, the problem of unbalanced category samples is well solved, the vehicle appearance image segmentation can be accurately completed, and the method can obtain higher accuracy on the vehicle appearance segmentation under the condition of low video memory occupation.

摘要：【中文】本发明公开了一种智能交通违法审核的图像筛选算法，涉及智能交通违法审核领域，解决了目前在违法审核图片筛选精度低，影响后续模块的处理精度的弊端，其技术方案要点是获取若干张图片，对获取的图片均进行语义分割，对分割结果不为全黑的图片形成轮廓并进行填充；对填充后的结果图一一进行比较，判断相比较的两张图片是否相同，对两张图的ID进行记录；根据每张图片被记录的次数进行排序、筛选判断出全景图及特写图；分别将语义分割时全黑的图片以及筛选获得的全景图及特写图进行车辆检测，再次进行判断筛选，本发明的一种智能交通违法审核的图像筛选算法，能够高精度的进行审核图片的全景图和特写图的区分，为后续模块提供高精度的支持。 【EN】The invention discloses an image screening algorithm for intelligent traffic violation audit, which relates to the field of intelligent traffic violation audit and solves the problems that the screening precision of the current illegal audit pictures is low and the processing precision of subsequent modules is influenced; comparing the filled result pictures one by one, judging whether the two compared pictures are the same or not, and recording the IDs of the two pictures; sorting and screening to judge a panoramic picture and a close-up picture according to the recorded times of each picture; the intelligent traffic violation checking image screening algorithm can accurately distinguish the panoramic image and the close-up image of the checked image, and provides high-accuracy support for subsequent modules.

摘要：【中文】本发明公开了一种基于车牌分类和LSTM的车牌识别方法，涉及车辆识别领域，解决了传统车辆识别算法复杂、精准度低，识别速度慢，其技术方案要点是获取待识别的车辆不同尺寸大小的图片，并对图像进行预处理和矫正调整；将调整后的图片输入分类网络中对车辆的车牌图片进行分类，根据图片中车牌类别分为单行车牌和双行车牌；将分类后的图片送入卷积神经网络，采样获得所需尺寸的特征图；若为双行车牌则进行分割并合并以获得新的特征图；将获得的特征图送入LSTM网络进行字符识别及预测，取概率最大的结果作为输出，检测获得整个车牌的字符，本发明的一种基于车牌分类和LSTM的车牌识别方法，实现了车牌的自动识别，实现端对端的识别，识别简便效果优良。 【EN】The invention discloses a license plate recognition method based on license plate classification and LSTM, which relates to the field of vehicle recognition and solves the problems of complex algorithm, low accuracy and low recognition speed of the traditional vehicle recognition; inputting the adjusted pictures into a classification network to classify the license plate pictures of the vehicles, and classifying the license plate pictures into single-row license plates and double-row license plates according to the license plate types in the pictures; sending the classified pictures into a convolutional neural network, and sampling to obtain a feature map with a required size; if the license plate is a double-row license plate, segmenting and combining the license plate to obtain a new feature map; the obtained feature map is sent to an LSTM network for character recognition and prediction, the result with the largest probability is taken as output, and the characters of the whole license plate are obtained through detection.

摘要：【中文】一种交通违法行为后台审核方法，包括步骤：获取目标车辆的疑似违法证据图像，标记所述疑似违法证据图像的至少一种疑似违法类型；基于所述疑似违法类型，从所述疑似违法证据图像中提取相应的违法要素信息；基于所述疑似违法类型，从预设违法取证标准数据库中获取标准违法要素信息；基于所述标准违法要素信息，根据从所述疑似违法证据图像中提取的违法要素信息，判别目标车辆是否违法；输出违法判别结果。本发明用机审代替人审，实现了对违法车辆图像证据的全天候研判，有效缓解了审核图片积压、废弃问题，减少了错判、漏判。 【EN】A background auditing method for traffic violation behaviors comprises the following steps: the method comprises the steps of obtaining a suspected illegal evidence image of a target vehicle, and marking at least one suspected illegal type of the suspected illegal evidence image; extracting corresponding illegal element information from the suspected illegal evidence image based on the suspected illegal type; acquiring standard illegal element information from a preset illegal evidence-taking standard database based on the suspected illegal type; judging whether a target vehicle is illegal or not according to illegal factor information extracted from the suspected illegal evidence image on the basis of the standard illegal factor information; and outputting the illegal judgment result. The invention replaces human review with machine review, realizes all-weather study and judgment of illegal vehicle image evidence, effectively relieves overstock and abandonment problems of the examined pictures, and reduces erroneous judgment and missed judgment.

摘要：【中文】本发明的目的是提供一种手写汉字图像恢复书写轨迹的方法及设备，本发明把静态手写汉字图片恢复出能够反应全部书写轨迹的序列图片问题转换为视频序列预测问题，使用笔画逐渐较少的序列图片训练deep RBM替代deep ConvLSTM的MCnet模型，把单张静态手写汉字图片输入到训练好的MCnet模型得到轨迹序列图片，实现对于给定的静态手写汉字图片,可以恢复出能够反应全部书写轨迹的图片序列。本发明相较于传统OCR方法，采用深度学习技术的本专利方案，鲁棒性更好。本发明不仅能应用于比较规整的手写图片，对于连笔手写图片，也能取得很好效果。恢复出的能够反应全部书写轨迹的图片序列可以应用于手写汉字的识别。 【EN】The invention aims to provide a method and equipment for restoring writing tracks of handwritten Chinese character images, which convert the problem that static handwritten Chinese character images restore sequence images capable of reflecting all writing tracks into the problem of video sequence prediction, train deep RBM to replace an MCnet model of deep ConvLSTM by using the sequence images with gradually fewer strokes, input a single static handwritten Chinese character image into the trained MCnet model to obtain a track sequence image, and realize that the image sequence capable of reflecting all writing tracks can be restored for the given static handwritten Chinese character image. Compared with the traditional OCR method, the method adopts the patent scheme of the deep learning technology, and has better robustness. The invention can be applied to regular hand-written pictures and can also obtain good effect on continuous hand-written pictures. The recovered picture sequence capable of reflecting all writing tracks can be applied to recognition of handwritten Chinese characters.

摘要：【中文】一种交强险保单车船税金额识别系统，包括第一目标检测模块，用于检测获取交强险保单上车船税金额数字在所述保单上的第一区域的粗略位置；目标标注模块，用于对于所述的第一区域中的车船税金额数字进行标注；第二目标检测模块，用于获取交强险保单中车船税金额数字在第一区域内的精确位置；字符识别模块，用于对车船税金额数字在进行格式统一、标注和增强处理后识别。 【EN】A kind of insurance policy vehicle and vessel tax amount recognition system of the insurance policy, including the first goal detection module, is used for detecting and obtaining the rough position of the vehicle and vessel tax amount figure on the insurance policy of the; the target marking module is used for marking the vehicle and ship tax amount figures in the first area; the second target detection module is used for acquiring the accurate position of the vehicle and ship tax amount number in the first area in the insurance policy of the traffic enforcement; and the character recognition module is used for recognizing the vehicle and ship tax amount figures after unified format, labeling and enhancement processing.

摘要：【中文】一种银行报表OCR识别方法，通过OCR获得银行报表的图像，对于银行报表的识别方法包括的步骤有：获得银行报表模板；检测银行报表图像的边界线段；对银行报表线段进行排除、融合、筛除处理后确定银行报表边界；对银行报表图像作仿射变换；确定银行报表上需识别区域并且对该区域进行识别。 【EN】A bank report OCR recognition method obtains an image of a bank report through OCR, and the recognition method of the bank report comprises the following steps: acquiring a bank report template; detecting a boundary line segment of a bank report image; removing, fusing and screening the bank report line segments, and determining the boundary of the bank report; performing affine transformation on the bank report image; and determining an area to be identified on the bank statement and identifying the area.

摘要：【中文】本申请涉及一种车辆外观检测方法、装置、计算机设备和存储介质。所述方法包括：将车辆拍摄图片输入至整车检测模型，得到车辆拍摄图片中的目标车辆图片；将目标车辆图片输入至语义分割模型，得到分割结果；分割结果中标注了至少一个车辆部件对应的待检测区域；根据各车辆部件对应的待检测区域及预设的检测条件，确定检测结果。采用本方法由于通过整车检测模型获取目标车辆图片并识别该目标车辆图片内的车辆部件，可以排除车辆拍摄图片中无关信息的干扰，提升了后续车辆部件识别过程的识别效率及识别准确性，进而可以使得检测结果更加准确，也提升了检测效率。 【EN】The application relates to a vehicle appearance detection method and device, a computer device and a storage medium. The method comprises the following steps: inputting the vehicle shooting picture into a whole vehicle detection model to obtain a target vehicle picture in the vehicle shooting picture; inputting the target vehicle picture into a semantic segmentation model to obtain a segmentation result; labeling the region to be detected corresponding to at least one vehicle part in the segmentation result; and determining a detection result according to the to-be-detected area corresponding to each vehicle component and a preset detection condition. By adopting the method, the target vehicle picture is obtained through the whole vehicle detection model and the vehicle part in the target vehicle picture is identified, so that the interference of irrelevant information in the vehicle shooting picture can be eliminated, the identification efficiency and the identification accuracy of the subsequent vehicle part identification process are improved, the detection result is more accurate, and the detection efficiency is also improved.

摘要：【中文】本申请公开了保险单的日期有效性判断方法、装置、设备及存储介质。保险单的日期有效性判断方法，通过获取保险单的图像中的起始日期和结束日期；基于深度学习的字符识别模型，获取针对起始日期的第一识别字符串和针对结束日期的第二识别字符串；根据第一识别字符串中连续的空白符的数量确定两个位置，并根据第二识别字符串中连续的空白符的数量确定另两个位置；根据两个位置划分第一识别字符串，并根据另两个位置划分第二识别字符串；根据当前日期，以及划分后的第一识别字符串和划分后的第二识别字符串，判断保险单的日期是否具备有效性；还公开了保险单的日期有效性判断装置、设备及存储介质；可提升保险单的日期的检查效率和检查结果的准确性。 【EN】The application discloses a date validity judgment method, a date validity judgment device, date validity judgment equipment and a storage medium of an insurance policy. The date validity judging method of the insurance policy comprises the steps of obtaining a starting date and an ending date in an image of the insurance policy; acquiring a first recognition character string for a start date and a second recognition character string for an end date based on the deeply learned character recognition model; determining two positions according to the number of continuous blank characters in the first identification character string, and determining the other two positions according to the number of continuous blank characters in the second identification character string; dividing a first recognition character string according to two positions, and dividing a second recognition character string according to the other two positions; judging whether the date of the insurance policy has validity or not according to the current date, the divided first identification character string and the divided second identification character string; also discloses date validity judging device, equipment and storage medium of insurance policy; the date checking efficiency of the insurance policy and the accuracy of the checking result can be improved.

摘要：【中文】本申请公开了一种车辆压线检测方法、装置、移动终端及存储介质，涉及智能交通领域，解决了目前车辆压线违法人工检测精度低、效率低的弊端，其技术方案要点是：获取待检测图像，待检测图像包括目标车辆和车道线；确定目标车辆的两个可信轮胎点；其中，可信轮胎点为：可基于待检测图像直观获取的轮胎触地点；根据两个可信轮胎点，确定目标车辆的另外两个轮胎的触地点；根据两个可信轮胎点、另外两个轮胎的触地点以及车道线，确定目标车辆是否压线；本申请还公开了一种车辆压线装置、移动终端及存储介质；可以准确定位车辆轮胎位置，检测车辆是否压线，判断过程高效简单，可提升检测效率和检测结果的精确性。 【EN】The application discloses vehicle line ball detection method, device, mobile terminal and storage medium relates to the intelligent transportation field, has solved the drawback that present vehicle line ball violation manual work detects the precision low, inefficiency, and its technical scheme main points are: acquiring an image to be detected, wherein the image to be detected comprises a target vehicle and a lane line; determining two credible tire points of the target vehicle; wherein, the credible tire points are as follows: a tire touchdown point visually obtainable based on an image to be detected; determining the contact points of the other two tires of the target vehicle according to the two credible tire points; determining whether the target vehicle presses the line or not according to the two credible tire points, the contact points of the other two tires and the lane line; the application also discloses a vehicle line pressing device, a mobile terminal and a storage medium; the vehicle tire position can be accurately positioned, whether the vehicle is pressed is detected, the judgment process is efficient and simple, and the detection efficiency and the accuracy of the detection result can be improved.