Keras image loading vs opencv imread function

[omar.m...@gmail.com]

Hi All,

I'm trying to use a keras model in production within an opencv application

I've a very weird observation when i load the image using opencv imread and preprocess it, the output is different from using keras, i expected the output to be slightly different but it's 10% difference in accuracy

for file in files:
    srcfile = root + '/' + file
    img = cv2.imread(root + '/' + file)
    img = scipy.misc.imresize(img, (224, 224), 'bilinear')
    new_im = np.zeros((3,224,224),dtype = np.float32)
    new_im[0,:,:] = img[:,:,2]
    new_im[1,:,:] = img[:,:,1]
    new_im[2,:,:] = img[:,:,0]
    img = new_im.copy()
    
    #second way
    im = image.load_img(srcfile,target_size = (224,224))
    im = np.asarray(im,dtype=np.uint8)
    im = im.transpose((2,0,1))
    im = np.expand_dims(im, axis= 0)
    img = np.expand_dims(img,axis= 0)
    preds , n , m=  vgg.predict(img)
    preds2, n2,m2 =  vgg.predict(im)

after looping through all images the classification performance, is a little bit different, can anyone guide me to the reason?

classification performance is totally different

[omar.m...@gmail.com]

Thank you guys

I found the problem, it was in the resize funtcion, opencv uses bilinear interpolation and keras use nearest neighbor

that's why the preprocessing is different