bytes, cv2, Image数据

女爷i 2023-02-13 11:23 13阅读 0赞

### Q: 问题描述 ###

A: 不同的数据加载方式，通过base64编码后，利用tornado请求算法服务时，返回结果不同。

### 现象 ###

*  图片内存大小为96,518 字节，先看下不同加载方式获得的二进制流的大小

>>> import cv2
    >>> from io import BytesIO  #py3
    >>> from PIL import Image
    >>> img_path='500_ChsEj17LW0eAWNZcAARjScabz30768.jpg'
    >>> bytes = open(img_path, 'rb').read()
    >>> img_cv = cv2.imread(img_path)
    >>> img_pil = Image.open(img_path)
    >>> flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv)
    >>> pil_buffer = BytesIO()
    >>> img_pil.save(pil_buffer, format='JPEG')
    >>> bytes_pil = pil_buffer.getvalue()
    >>> len(bytes)
    96518
    >>> len(img_buff_cv)
    143421
    >>> len(img_buff_cv.tobytes())
    143421
    >>> len(img_buff_cv.tostring())
    143421
    >>> len(bytes_pil)
    80360

*  请求结果。其中请求数据采用base64编码，服务端解码后转换为opencv图片格式

解码端数据格式转换
        file_bytes = np.asarray(bytearray(img_file), dtype=np.uint8)
        img = cv2.imdecode(file_bytes, 1)
    请求数据结果
    #bytes的base64格式
        'car': [[30.56024169921875, 144.45347595214844, 573.21044921875, 413.3760986328125,         0.9995958209037781]],
        'person': [[197.00421142578125, 65.57933044433594, 295.8858947753906, 433.0203857421875, 0.9970766305923462]]
    #img_buff_cv的base64格式
        'car': [[82.89448547363281, 131.2141876220703, 571.5702514648438, 423.4444580078125, 0.9993151426315308]]
        'person': [[196.47152709960938, 65.93354797363281, 294.8117370605469, 432.09088134765625, 0.9969106316566467]]
    #img_buff_cv.tobytes的base64格式
        'car': [[82.89448547363281, 131.2141876220703, 571.5702514648438, 423.4444580078125, 0.9993151426315308]]
        'person': [[196.47152709960938, 65.93354797363281, 294.8117370605469, 432.09088134765625, 0.9969106316566467]]
    #img_buff_cv.tostring的base64格式
        'car': [[82.89448547363281, 131.2141876220703, 571.5702514648438, 423.4444580078125, 0.9993151426315308]]
        'person': [[196.47152709960938, 65.93354797363281, 294.8117370605469, 432.09088134765625, 0.9969106316566467]]
    #bytes_pil的base64格式
        'car': [[54.25007629394531, 151.00857543945312, 552.50927734375, 412.3731994628906, 0.9993000030517578]]
        'person': [[193.79055786132812, 66.72914123535156, 295.7754821777344, 428.609130859375, 0.9984044432640076]]

*  观察检测结果发现(单个样本，不确定是否有通用性)，二进制直接编码后的请求结果更好些。现在减少opencv数据格式的内存，重新观察结果。

>>> flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,100])
    >>> len(img_buff_cv)
    221349
    >>> flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [2,100])
    >>> len(img_buff_cv)
    116289
    >>> flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,87])
    >>> len(img_buff_cv)
    97963
    >>> flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,80])
    >>> len(img_buff_cv)
    86785
    
    #flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,100])
    	'car': [[30.692169189453125, 144.15696716308594, 573.0172119140625, 413.498291015625, 0.9995959401130676]]
    	'person': [[196.8931884765625, 65.76217651367188, 295.91668701171875, 433.08538818359375, 0.9969993829727173]]
    #flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [2,100])
    	'car': [[82.89448547363281, 131.2141876220703, 571.5702514648438, 423.4444580078125, 0.9993151426315308]]
    	'person': [[196.47152709960938, 65.93354797363281, 294.8117370605469, 432.09088134765625, 0.9969106316566467]]
    #flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,87])
    	'car': [[32.0150146484375, 144.37466430664062, 573.1845703125, 413.5013732910156, 0.999593198299408]]
    	'person': [[196.95909118652344, 65.48707580566406, 295.9208984375, 433.371826171875, 0.9970276951789856]]
    #flag_cvt, img_buff_cv = cv2.imencode('.jpg', img_cv, [1,80])
    	'car': [[23.19580078125, 142.8824462890625, 576.1643676757812, 413.4249267578125, 0.9995841383934021]]
    	'person': [[197.58522033691406, 59.73590087890625, 299.53619384765625, 436.0, 0.9977819323539734]]

*  结论，使用opencv数据转换为二进制数据时，图像编码质量会影响最终的算法效果。

### 参考文献： ###

1.  [Convert PIL or OpenCV Image to Bytes without Saving to Disk][]
2.  [How to read image from in memory buffer (StringIO) or from url with opencv python library][How to read image from in memory buffer _StringIO_ or from url with opencv python library]
3.  [PIL.Image与Base64 String的互相转换][PIL.Image_Base64 String]

[Convert PIL or OpenCV Image to Bytes without Saving to Disk]: https://jdhao.github.io/2019/07/06/python_opencv_pil_image_to_bytes/
[How to read image from in memory buffer _StringIO_ or from url with opencv python library]: https://stackoverflow.com/questions/13329445/how-to-read-image-from-in-memory-buffer-stringio-or-from-url-with-opencv-pytho
[PIL.Image_Base64 String]: https://www.jianshu.com/p/2ff8e6f98257