Files
immich/server/src/services/ocr.service.ts
Kang 02b29046b3 feat: ocr (#18836)
* feat: add OCR functionality and related configurations

* chore: update labeler configuration for machine learning files

* feat(i18n): enhance OCR model descriptions and add orientation classification and unwarping features

* chore: update Dockerfile to include ccache for improved build performance

* feat(ocr): enhance OCR model configuration with orientation classification and unwarping options, update PaddleOCR integration, and improve response structure

* refactor(ocr): remove OCR_CLEANUP job from enum and type definitions

* refactor(ocr): remove obsolete OCR entity and migration files, and update asset job status and schema to accommodate new OCR table structure

* refactor(ocr): update OCR schema and response structure to use individual coordinates instead of bounding box, and adjust related service and repository files

* feat: enhance OCR configuration and functionality

- Updated OCR settings to include minimum detection box score, minimum detection score, and minimum recognition score.
- Refactored PaddleOCRecognizer to utilize new scoring parameters.
- Introduced new database tables for asset OCR data and search functionality.
- Modified related services and repositories to support the new OCR features.
- Updated translations for improved clarity in settings UI.

* sql changes

* use rapidocr

* change dto

* update web

* update lock

* update api

* store positions as normalized floats

* match column order in db

* update admin ui settings descriptions

fix max resolution key

set min threshold to 0.1

fix bind

* apply config correctly, adjust defaults

* unnecessary model type

* unnecessary sources

* fix(ocr): switch RapidOCR lang type from LangDet to LangRec

* fix(ocr): expose lang_type (LangRec.CH) and font_path on OcrOptions for RapidOCR

* fix(ocr): make OCR text search case- and accent-insensitive using ILIKE + unaccent

* fix(ocr): add OCR search fields

* fix: Add OCR database migration and update ML prediction logic.

* trigrams are already case insensitive

* add tests

* format

* update migrations

* wrong uuid function

* linting

* maybe fix medium tests

* formatting

* fix weblate check

* openapi

* sql

* minor fixes

* maybe fix medium tests part 2

* passing medium tests

* format web

* readd sql

* format dart

* disabled in e2e

* chore: translation ordering

---------

Co-authored-by: mertalev <101130780+mertalev@users.noreply.github.com>
Co-authored-by: Alex Tran <alex.tran1502@gmail.com>
2025-10-27 14:09:55 +00:00

87 lines
2.7 KiB
TypeScript

import { Injectable } from '@nestjs/common';
import { JOBS_ASSET_PAGINATION_SIZE } from 'src/constants';
import { OnJob } from 'src/decorators';
import { AssetVisibility, JobName, JobStatus, QueueName } from 'src/enum';
import { OCR } from 'src/repositories/machine-learning.repository';
import { BaseService } from 'src/services/base.service';
import { JobItem, JobOf } from 'src/types';
import { isOcrEnabled } from 'src/utils/misc';
@Injectable()
export class OcrService extends BaseService {
@OnJob({ name: JobName.OcrQueueAll, queue: QueueName.Ocr })
async handleQueueOcr({ force }: JobOf<JobName.OcrQueueAll>): Promise<JobStatus> {
const { machineLearning } = await this.getConfig({ withCache: false });
if (!isOcrEnabled(machineLearning)) {
return JobStatus.Skipped;
}
if (force) {
await this.ocrRepository.deleteAll();
}
let jobs: JobItem[] = [];
const assets = this.assetJobRepository.streamForOcrJob(force);
for await (const asset of assets) {
jobs.push({ name: JobName.Ocr, data: { id: asset.id } });
if (jobs.length >= JOBS_ASSET_PAGINATION_SIZE) {
await this.jobRepository.queueAll(jobs);
jobs = [];
}
}
await this.jobRepository.queueAll(jobs);
return JobStatus.Success;
}
@OnJob({ name: JobName.Ocr, queue: QueueName.Ocr })
async handleOcr({ id }: JobOf<JobName.Ocr>): Promise<JobStatus> {
const { machineLearning } = await this.getConfig({ withCache: true });
if (!isOcrEnabled(machineLearning)) {
return JobStatus.Skipped;
}
const asset = await this.assetJobRepository.getForOcr(id);
if (!asset || !asset.previewFile) {
return JobStatus.Failed;
}
if (asset.visibility === AssetVisibility.Hidden) {
return JobStatus.Skipped;
}
const ocrResults = await this.machineLearningRepository.ocr(asset.previewFile, machineLearning.ocr);
await this.ocrRepository.upsert(id, this.parseOcrResults(id, ocrResults));
await this.assetRepository.upsertJobStatus({ assetId: id, ocrAt: new Date() });
this.logger.debug(`Processed ${ocrResults.text.length} OCR result(s) for ${id}`);
return JobStatus.Success;
}
private parseOcrResults(id: string, { box, boxScore, text, textScore }: OCR) {
const ocrDataList = [];
for (let i = 0; i < text.length; i++) {
const boxOffset = i * 8;
ocrDataList.push({
assetId: id,
x1: box[boxOffset],
y1: box[boxOffset + 1],
x2: box[boxOffset + 2],
y2: box[boxOffset + 3],
x3: box[boxOffset + 4],
y3: box[boxOffset + 5],
x4: box[boxOffset + 6],
y4: box[boxOffset + 7],
boxScore: boxScore[i],
textScore: textScore[i],
text: text[i],
});
}
return ocrDataList;
}
}