在批量检测图片时控制台可能会出现类似下面的日志Corrupt JPEG data: 53 extraneous bytes before marker 0xd9 Corrupt JPEG data: premature end of data segment这类日志通常不是 Python 主动抛出的异常而是 OpenCV 底层 JPEG 解码库输出到stderr的警告。很多情况下cv2.imread()仍然会返回图像但图像数据已经存在截断、尾部异常字节或写入不完整的问题。对于目标检测任务这类图片可能导致检测框异常、漏检、置信度波动甚至影响批量处理稳定性。产生原因常见原因包括图片采集或网络传输未完成文件只保存了一部分。写文件过程中程序退出、磁盘异常或进程被终止。JPEG 文件尾部存在多余字节触发extraneous bytes before marker 0xd9。JPEG 数据段提前结束触发premature end of data segment。Windows 环境下直接使用cv2.imread(path)读取中文路径可能把正常图片误判为读取失败。其中第 5 点很重要如果图片路径包含中文不能简单使用cv2.imread()判断图片是否损坏。更稳妥的方式是datanp.fromfile(path,dtypenp.uint8)imgcv2.imdecode(data,cv2.IMREAD_COLOR)这种方式由 Python 负责读取文件路径能够正确处理中文路径再交给 OpenCV 解码图片内容。处理思路推荐按下面顺序处理先扫描图片目录只生成坏图报告不修改文件。人工检查bad_images.csv确认是否确实是坏图。第一次处理建议移动到隔离目录不要直接删除。确认无误后再选择是否删除。如果图片旁边有同名 LabelMe JSON可以使用--with-json一起移动或删除。完整代码将下面代码保存为find_bad_images.pyimportargparseimportcsvimportosimportshutilimportsubprocessimportsysfrompathlibimportPath IMAGE_EXTS{.jpg,.jpeg,.png,.bmp,.tif,.tiff,.webp}JPEG_WARNING_PATTERNS(corrupt jpeg data,premature end of data segment,extraneous bytes before marker,invalid sos parameters,bad huffman code,unsupported marker type,)CHECK_CODEr import sys import cv2 import numpy as np data np.fromfile(sys.argv[1], dtypenp.uint8) if data.size 0: print(image file is empty or unreadable, filesys.stderr) sys.exit(2) img cv2.imdecode(data, cv2.IMREAD_COLOR) if img is None: print(cv2.imdecode returned None, filesys.stderr) sys.exit(2) sys.exit(0) defparse_args():parserargparse.ArgumentParser(descriptionFind JPEG/images that trigger OpenCV decode warnings.)parser.add_argument(--image-dir,default../Car,helpDirectory to scan.)parser.add_argument(--recursive,actionstore_true,helpScan image-dir recursively.)parser.add_argument(--report,defaultbad_images.csv,helpCSV report path.)parser.add_argument(--move-bad,defaultNone,helpMove bad images to this directory instead of deleting.)parser.add_argument(--delete,actionstore_true,helpDelete bad images. Use only after checking the report.)parser.add_argument(--with-json,actionstore_true,helpAlso move/delete same-stem LabelMe json files.)parser.add_argument(--any-stderr,actionstore_true,helpTreat any decoder stderr output as bad.)returnparser.parse_args()defiter_images(image_dir:Path,recursive:bool):pattern**/*ifrecursiveelse*forpathinsorted(image_dir.glob(pattern)):ifpath.is_file()andpath.suffix.lower()inIMAGE_EXTS:yieldpathdefcheck_image(path:Path,any_stderr:bool):procsubprocess.run([sys.executable,-c,CHECK_CODE,str(path)],stdoutsubprocess.PIPE,stderrsubprocess.PIPE,textTrue,encodingutf-8,errorsreplace,)stderrproc.stderr.strip()stderr_lowerstderr.lower()has_known_warningany(patterninstderr_lowerforpatterninJPEG_WARNING_PATTERNS)ifproc.returncode!0:returnFalse,stderrorfdecode failed with return code{proc.returncode}ifhas_known_warning:returnFalse,stderrifany_stderrandstderr:returnFalse,stderrreturnTrue,defrelated_files(image_path:Path,include_json:bool):paths[image_path]ifinclude_json:json_pathimage_path.with_suffix(.json)ifjson_path.exists():paths.append(json_path)returnpathsdefmove_files(paths,source_root:Path,target_root:Path):moved[]forpathinpaths:relative_pathpath.relative_to(source_root)target_pathtarget_root/relative_path target_path.parent.mkdir(parentsTrue,exist_okTrue)shutil.move(str(path),str(target_path))moved.append(str(target_path))returnmoveddefdelete_files(paths):deleted[]forpathinpaths:path.unlink()deleted.append(str(path))returndeleteddefmain():argsparse_args()image_dirPath(args.image_dir).resolve()ifnotimage_dir.exists():raiseFileNotFoundError(fimage-dir does not exist:{image_dir})ifargs.deleteandargs.move_bad:raiseValueError(--delete and --move-bad cannot be used together)report_pathPath(args.report)bad_rows[]total0forimage_pathiniter_images(image_dir,args.recursive):total1ok,reasoncheck_image(image_path,args.any_stderr)ifok:continueactionreportaffected[]filesrelated_files(image_path,args.with_json)ifargs.move_bad:actionmoveaffectedmove_files(files,image_dir,Path(args.move_bad).resolve())elifargs.delete:actiondeleteaffecteddelete_files(files)bad_rows.append([str(image_path),reason.replace(\n, | ),action,;.join(affected)])print(fBAD:{image_path})print(f reason:{reason})ifaffected:print(f{action}:{affected})withreport_path.open(w,newline,encodingutf-8)asreport_file:writercsv.writer(report_file)writer.writerow([image_path,reason,action,affected_files])writer.writerows(bad_rows)print()print(fScanned images:{total})print(fBad images:{len(bad_rows)})print(fReport:{report_path.resolve()})ifbad_rowsandnotargs.deleteandnotargs.move_bad:print(No files were changed. Re-run with --move-bad bad_images or --delete after checking the report.)if__name____main__:main()使用方法只扫描不修改任何图片python find_bad_images.py --image-dirD:\数据集\车辆图片--reportbad_images.csv递归扫描子目录python find_bad_images.py --image-dirD:\数据集\车辆图片--recursive--reportbad_images.csv推荐先移动坏图到隔离目录python find_bad_images.py --image-dirD:\数据集\车辆图片--move-bad bad_images --with-json确认坏图无价值后直接删除python find_bad_images.py --image-dirD:\数据集\车辆图片--delete--with-json如果想把所有 OpenCV 解码stderr输出都当作异常处理可以加python find_bad_images.py --image-dirD:\数据集\车辆图片--any-stderr--reportbad_images.csv报告字段说明bad_images.csv包含以下字段字段说明image_path被判定异常的图片路径reasonOpenCV/JPEG 解码输出的异常原因action当前执行动作可能是report、move或deleteaffected_files被移动或删除的文件路径注意事项如果路径包含中文必须使用本文代码里的np.fromfile cv2.imdecode不要直接用cv2.imread(path)做坏图判断。第一次处理建议使用--move-bad不要直接--delete。如果已经生成过旧版报告应删除旧的bad_images.csv后重新扫描。--with-json适合 LabelMe 数据集会同步处理同名.json标注文件。如果图片来自摄像头、网络请求或异步写盘流程应同时检查上游写文件逻辑避免还没写完就进入检测流程。