The image output by PiNoIR Camera Module on an old Raspbeery Pi 1 MODEL B+ is sent to Classify ImageNet classes with ResNet50.
model = ResNet50(weights='imagenet')
reads ResNet50. It takes about seven minutes by Raspbeery Pi 1 MODEL B+...
while True:
with picamera.array.PiRGBArray(camera) as stream:
camera.capture(stream, 'bgr')
image = stream.array
if WINDOW_WIDTH != ORG_WIDTH:
image = cv2.resize(image, (WINDOW_WIDTH, WINDOW_HEIGHT), interpolation = cv2.INTER_AREA)
cv2.imshow(windowName, image)
obtains BGR ordered OpenCV structured numpy array from PiCamera and show as an image with OpenCV imgshow() Ref. 4.4. Capturing to an OpenCV object
The obtained image is resized with OpenCV resize() to the window size.
key = cv2.waitKey(12)
#press Esc(27) to quit, press c(99) to classify
if key==27:
break
elif key==99:
print('classifying image...')
t2 = time.clock()
image = image[: ,w_offset:w_offset + ORG_HEIGHT, :]
An inference is only executed at the time when the key c is pressed, because the inference takes about 50 seconds... At first, an obtained image is resized to ORG_HEIGHT x ORG_HEIGHT x 3 as the same center as ORG_WIDGH by offsetting (ORG_WIDTH - ORG_HEIGHT)/2 pixels.
image = cv2.resize(image, (224, 224), interpolation = cv2.INTER_AREA)
x = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32)
array_to_img(x).save('classified.jpg')
resize ORG_HEIGHT x ORG_HEIGHT x 3 to 224 x 224 x 3. OpenCV cvColor() transforms a BGR ordered image with uint type to a ResNet50 required RGB ordered one with float type. The image sent to ResNet50 is stored with Keras array_to_img and PIL save() as classified.jpg.
x = np.expand_dims(x, axis=0)
x = preprocess_input(x)
preds = model.predict(x)
print('Predicted:', decode_predictions(preds))
t3 = time.clock()
print('inference : %.3f s' % (t3 - t2))
The later is same as Classify ImageNet classes with ResNet50.
How an input image is shown and the inference is output like this:
The banana is 93% likely banana.
