世界人工智能大赛 Top1 方案!手写体 OCR 识别

 Datawhale干货 

作者:王浩,结行科技算法工程师

参加了“世界人工智能创新大赛”——手写体 OCR 识别竞赛(任务一),取得了Top1的成绩。队伍随机组的,有人找我我就加了进来,这是我第一次做OCR相关的项目,所以随意起了个名字。下面通过这篇文章来详细介绍我们的方案。

478af0112b51351aae8199f4f2190cb3.png

实践背景

赛题背景

银行日常业务中涉及到各类凭证的识别录入,例如身份证录入、支票录入、对账单录入等。以往的录入方式主要是以人工录入为主,效率较低,人力成本较高。近几年来,OCR相关技术以其自动执行、人为干预较少等特点正逐步替代传统的人工录入方式。但OCR技术在实际应用中也存在一些问题,在各类凭证字段的识别中,手写体由于其字体差异性大、字数不固定、语义关联性较低、凭证背景干扰等原因,导致OCR识别率准确率不高,需要大量人工校正,对日常的银行录入业务造成了一定的影响。

赛题地址:http://ailab.aiwin.org.cn/competitions/65

赛题任务

本次赛题将提供手写体图像切片数据集,数据集从真实业务场景中,经过切片脱敏得到,参赛队伍通过识别技术,获得对应的识别结果。即:

  • 输入:手写体图像切片数据集

  • 输出:对应的识别结果

本任务提供开放可下载的训练集及测试集,允许线下建模或线上提供 Notebook 环境及 Terminal 容器环境(脱网)建模,输出识别结果完成赛题。

赛题数据

A. 数据规模和内容覆盖

56cea6687be090e0a4ea55ef9ad38cc0.png

B.数据示例

原始手写体图像共分为三类,分别涉及银行名称、年月日、金额三大类,分别示意如下:

4cfda26453767dec3b2d436540d3c6b8.png

相应图片切片中可能混杂有一定量的干扰信息,分别示例如下:

cd77146fdad3c8b41f455b2992a6b784.png

识别结果 JSON 在训练集中的格式如下:

json 文件内容规范:
{
 "image1": "陆万捌千零贰拾伍元整",
 "image2": "付经管院工资",
 "image3": "",
 ...
}

实践方案

通过在网上查阅资料,得知OCR比赛最常用的模型是CRNN+CTC。所以我最开始也是采用这个方案。

67d53ddb400059fa88f885b8a9a8f35b.png

上图是我找到的资料,有好多个版本。因为是第一次做OCR的项目,所以我优先选择有数据集的项目,这样可以快速的了解模型的输入输出。

所以我选择的第一个Attention_ocr.pytorch-master.zip,从名字上可以看出这个是加入注意力机制,感觉效果会好一些。

构建数据集

下图是Attention_ocr.pytorch-master.zip自带的数据集截图,从截图上可以看出,数据的格式:“图片路径+空格+标签”。我们也需要按照这样的格式构建数据集。

53061bb93c916ab64926814ebac5f02c.png

新建makedata.py文件,插入下面的代码。

import os
import json
#官方给的数据集
image_path_amount = "./data/train/amount/images" 
image_path_date = "./data/train/date/images"
#增强数据集
image_path_test='./data/gan_test_15000/images/0'
image_path_train='./data/gan_train_15500_0/images/0'
amount_list = os.listdir(image_path_amount)
amount_list = os.listdir(image_path_amount)

new_amount_list = []
for filename in amount_list:
    new_amount_list.append(image_path_amount + "/" + filename)

date_list = os.listdir(image_path_date)
new_date_list = []
for filename in date_list:
    new_date_list.append(image_path_date + "/" + filename)
new_test_list = []
for filename in amount_list:
    new_test_list.append(image_path_amount + "/" + filename)

new_train_list = []
for filename in amount_list:
    new_train_list.append(image_path_amount + "/" + filename)
05-23 16:00:49.311 25132-25512 A0c0d0/JSAPP I result {"code":200,"msg":"操作成功","data":[{"id":1,"name":"华为Mate 60 Pro","model":"ALN-AL80","imageUrl":"https://img.alicdn.com/imgextra/i3/2213047591560/O1CN01QNx6eB1NOWu3oeUfn_!!2213047591560.jpg_q50.jpg_.webp","shortDescription":"麒麟9000S 卫星通信","price":6999.00,"subsidizedPrice":6499.00,"sales":15000,"rating":4.90,"stock":500,"isAvailable":0,"brand":"华为"},{"id":2,"name":"华为P60 Art","model":"NOH-AN80","imageUrl":"https://img.alicdn.com/imgextra/i4/2609173245/O1CN011h2L1I1ZqGCdlsYlJ_!!2609173245.jpg_q50.jpg_.webp","shortDescription":"超聚光XMAGE影像","price":7988.00,"subsidizedPrice":7588.00,"sales":12000,"rating":4.80,"stock":300,"isAvailable":0,"brand":"华为"},{"id":3,"name":"华为nova 12","model":"BAC-AL00","imageUrl":"https://picasso.alicdn.com/imgextra/O1CNA1bnbda21Vuba4TjLYB_!!2838892713-0-psf.jpg_.webp","shortDescription":"前置双摄人像","price":2999.00,"subsidizedPrice":2699.00,"sales":50000,"rating":4.70,"stock":1000,"isAvailable":0,"brand":"华为"},{"id":4,"name":"荣耀Magic6 Pro","model":"PGT-AN30","imageUrl":"https://img.alicdn.com/imgextra/i2/263726286/O1CN01ko93gt1wJ2hQAORAB_!!4611686018427380942-0-item_pic.jpg_.webp","shortDescription":"骁龙8 Gen3 青海湖电池","price":5699.00,"subsidizedPrice":5399.00,"sales":25000,"rating":4.80,"stock":600,"isAvailable":0,"brand":"荣耀"},{"id":5,"name":"荣耀X50 GT","model":"ANY-AN00","imageUrl":"https://img.alicdn.com/imgextra/i3/3851598352/O1CN01UHPfyv2BZGwPnHUlx_!!0-item_pic.jpg_q50.jpg_.webp","shortDescription":"1.5K护眼曲屏","price":1999.00,"subsidizedPrice":1799.00,"sales":80000,"rating":4.60,"stock":1500,"isAvailable":0,"brand":"荣耀"},{"id":6,"name":"荣耀90 Pro","model":"REA-AN20","imageUrl":"https://img.alicdn.com/imgextra/i3/1114511827/O1CN01r3ngr41PMof7MhFhY_!!4611686018427386323-0-item_pic.jpg_q50.jpg_.webp","shortDescription":"2亿像素写真相机","price":3299.00,"subsidizedPrice":2999.00,"sales":40000,"rating":4.70,"stock":800,"isAvailable":0,"brand":"荣耀"},{"id":7,"name":"荣耀Play7T Pro","model":"CMA-AN40","imageUrl":"https://img.alicdn.com/imgextra/i1/1907069314/O1CN019jd3kt2IfrnZ6J0AH_!!1907069314.jpg_q50.jpg_.webp","shortDescription":"40W快充 OLED屏","price":1799.00,"subsidizedPrice":1599.00,"sales":120000,"rating":4.50,"stock":2000,"isAvailable":0,"brand":"荣耀"},{"id":8,"name":"华为畅享70 Pro","model":"CMA-AN70","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"7000mAh长续航","price":1599.00,"subsidizedPrice":1399.00,"sales":150000,"rating":4.40,"stock":3000,"isAvailable":0,"brand":"华为"},{"id":9,"name":"华为Mate X5","model":"TET-AN80","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"折叠屏 玄武钢化","price":12999.00,"subsidizedPrice":12499.00,"sales":5000,"rating":4.90,"stock":100,"isAvailable":0,"brand":"华为"},{"id":10,"name":"荣耀Magic Vs2","model":"CMA-AN80","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"轻薄折叠屏","price":8999.00,"subsidizedPrice":8499.00,"sales":10000,"rating":4.70,"stock":300,"isAvailable":0,"brand":"荣耀"},{"id":11,"name":"iPhone 15 Pro Max","model":"A3108","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"A17 Pro 钛金属","price":9999.00,"subsidizedPrice":null,"sales":80000,"rating":4.80,"stock":200,"isAvailable":0,"brand":"苹果"},{"id":12,"name":"iPhone 15","model":"A3092","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"A16芯片 灵动岛","price":5999.00,"subsidizedPrice":null,"sales":150000,"rating":4.80,"stock":5000,"isAvailable":0,"brand":"苹果"},{"id":13,"name":"iPhone 15 Plus","model":"A3096","imageUrl":"https://android-api.oss-cn-beijing.aliyuncs.com/logo.jpg","shortDescription":"6.7英寸大屏","price":6999.00,"subsidizedPrice":null,"sales":80000,"rating":4.70,"stock":3000,"isAvailable":0,"brand":"苹果"},{"id":14,"name":"iPhone SE 4","model":"A2785","imageUrl":"https://android-api.oss-cn-beijing.aliyun 05-23 16:00:49.311 25132-25512 A0c0d0/JSAPP I result [object Object] 05-23 16:00:49.359 25132-25512 C03f00/ArkCompiler E [ArkRuntime Log] TypeError: is not callable 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log]Lifetime: 0.000000s 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log]Js-Engine: ark 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log]page: pages/Index.js 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log]Error message: is not callable 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log]Stacktrace: 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at forEachUpdateFunction (/devcloud/slavespace/usr1/081f8aba80800f0f0fcec015bd66c7e0/harmony_code/codearts_workspace/foundation/arkui/ace_engine/frameworks/bridge/declarative_frontend/engine/stateMgmt.js:4352:1) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at anonymous (entry|entry|1.0.0|src/main/ets/pages/Index.ts:144:13) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at updateFunc (/devcloud/slavespace/usr1/081f8aba80800f0f0fcec015bd66c7e0/harmony_code/codearts_workspace/foundation/arkui/ace_engine/frameworks/bridge/declarative_frontend/engine/stateMgmt.js:6812:1) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at UpdateElement (/devcloud/slavespace/usr1/081f8aba80800f0f0fcec015bd66c7e0/harmony_code/codearts_workspace/foundation/arkui/ace_engine/frameworks/bridge/declarative_frontend/engine/stateMgmt.js:6514:1) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at anonymous (/devcloud/slavespace/usr1/081f8aba80800f0f0fcec015bd66c7e0/harmony_code/codearts_workspace/foundation/arkui/ace_engine/frameworks/bridge/declarative_frontend/engine/stateMgmt.js:6741:1) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at updateDirtyElements (/devcloud/slavespace/usr1/081f8aba80800f0f0fcec015bd66c7e0/harmony_code/codearts_workspace/foundation/arkui/ace_engine/frameworks/bridge/declarative_frontend/engine/stateMgmt.js:6736:1) 05-23 16:00:49.361 25132-25512 C03900/Ace E [Engine Log] at rerender (entry|entry|1.0.0|src/main/ets/pages/Index.ts:152:9)
最新发布
05-24
### 关于 HarmonyOS 或 ArkCompiler 中 'is not callable' 的 TypeError 问题 在 HarmonyOS 和 ArkCompiler 开发环境中遇到 `TypeError: is not callable` 错误通常表明尝试调用的对象并非函数或可调用对象。这种错误可能由多种原因引起,例如变量名冲突、未正确定义的回调函数或者对 `null` 值进行了不当操作。 #### 可能的原因分析 1. **变量覆盖了内置方法或函数** 如果开发者无意间将某些全局名称重新赋值为其他类型的值(如 `null`),可能会导致该错误。例如,在 JavaScript 环境下,如果重写了 `Array.prototype.map` 方法并将其设置为 `null`,那么后续对该方法的调用会失败[^4]。 2. **异步编程中的上下文丢失** 在使用箭头函数或其他形式绑定事件处理器时,如果没有正确处理 this 上下文,则可能导致原本期望作为函数执行的内容变成了 undefined 或 null。这种情况常见于 React 组件生命周期方法以及 Vue.js 自定义指令场景中[^5]。 3. **API 返回值异常** 当从外部服务获取数据时,若返回的结果不符合预期结构——比如应该是一个函数却实际接收到的是 null ——也会引发此类错误消息提示[^6]。 4. **Null 安全性问题** 类似于 MySQL 数据库设计里提到允许字段为空 (NULL)[^2], 在程序逻辑层面也需要考虑如何优雅地应对可能出现的 Null 情况。如果不小心试图直接调用了被赋予 Null 的属性或方法就会抛出类似的运行期错误。 #### 解决方案建议 针对以上几种可能性提供相应的解决方案如下: - #### 验证目标是否确实具备可调用特性 在每次准备调用前先验证其类型是否属于 Function 。可以利用 typeof 运算符来进行初步判断。 ```javascript const maybeFunction = someCondition ? () => {} : null; if(typeof maybeFunction === 'function') { maybeFunction(); } ``` - #### 使用 Optional Chaining 提高健壮性 ES2020 引入了一个新功能叫做 optional chaining (`?.`) ,它可以帮助我们更简洁安全地访问深层嵌套对象而不用担心中间环节存在 null/undefined 导致崩溃风险。 ```javascript let result = possiblyUndefinedObject ?. methodToCall(); // 不会报错即使object不存在 ``` - #### 明确区分命名空间内的同名实体 尽量避免自定义标识符与框架内部保留字发生碰撞;另外通过模块化组织代码也能有效减少这类隐患的发生几率。 - #### 调试工具辅助定位具体位置 利用断点调试技术逐步跟踪哪一部分代码最终触发了这个特定异常,并针对性修复源头问题所在之处。 ```javascript try{ potentiallyProblematicCode(); }catch(e){ console.error('Error occurred:', e); } ``` ### 总结 通过对 HarmonyOS / ArkCompiler 平台开发过程中常见的 “not callable” 类型错误现象剖析可知,这往往涉及到基本概念理解偏差或者是编码习惯上的疏漏所致。采取适当预防措施加上良好实践能够显著降低遭遇这些问题的概率。
评论 2
添加红包

请填写红包祝福语或标题

红包个数最小为10个

红包金额最低5元

当前余额3.43前往充值 >
需支付:10.00
成就一亿技术人!
领取后你会自动成为博主和红包主的粉丝 规则
hope_wisdom
发出的红包
实付
使用余额支付
点击重新获取
扫码支付
钱包余额 0

抵扣说明:

1.余额是钱包充值的虚拟货币,按照1:1的比例进行支付金额的抵扣。
2.余额无法直接购买下载,可以购买VIP、付费专栏及课程。

余额充值