def detect(self, original_image):
self.output_width, self.output_height = original_image.shape[0:2]
start_time = time.time()
image = cv2.resize(original_image,
(self.input_width, self.input_height))
#image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
results = self.detect_objects(image)
elapsed_ms = (time.time() - start_time) * 1000
fps = 1 / elapsed_ms * 1000
print("Estimated frames per second : {0:.2f} Inference time: {1:.2f}".
format(fps, elapsed_ms))
def _to_original_scale(boxes):
minmax_boxes = to_minmax(boxes)
minmax_boxes[:, 0] *= self.output_width
minmax_boxes[:, 2] *= self.output_width
minmax_boxes[:, 1] *= self.output_height
minmax_boxes[:, 3] *= self.output_height
return minmax_boxes.astype(np.int)
boxes, probs = self.run(results)
print(boxes)
if len(boxes) > 0:
boxes = _to_original_scale(boxes)
original_image = draw_boxes(original_image, boxes, probs,
self.labels)
return cv2.imencode('.jpg', original_image)[1].tobytes()
def parse(self, original_image, tensor):
#start_time = time.time()
#elapsed_ms = (time.time() - start_time) * 1000
#fps = 1 / elapsed_ms*1000
#print("Estimated frames per second : {0:.2f} Inference time: {1:.2f}".format(fps, elapsed_ms))
boxes, probs = self.run(tensor)
def _to_original_scale(boxes):
minmax_boxes = to_minmax(boxes)
minmax_boxes[:, 0] *= 224
minmax_boxes[:, 2] *= 224
minmax_boxes[:, 1] *= 224
minmax_boxes[:, 3] *= 224
return minmax_boxes.astype(np.int)
if len(boxes) > 0:
boxes = _to_original_scale(boxes)
#print(boxes)
original_image = draw_boxes(original_image, boxes, probs,
self.labels)
return original_image
def detect(self, original_image):
self.output_height, self.output_width = original_image.shape[0:2]
start_time = time.time()
results = self.detect_objects(original_image)
elapsed_ms = (time.time() - start_time) * 1000
fps = 1 / elapsed_ms * 1000
print("Estimated frames per second : {0:.2f} Inference time: {1:.2f}".
format(fps, elapsed_ms))
def _to_original_scale(boxes):
minmax_boxes = to_minmax(boxes)
minmax_boxes[:, 0] *= self.output_width
minmax_boxes[:, 2] *= self.output_width
minmax_boxes[:, 1] *= self.output_height
minmax_boxes[:, 3] *= self.output_height
return minmax_boxes.astype(np.int)
boxes, probs = self.run(results)
print(boxes)
if len(boxes) > 0:
boxes = _to_original_scale(boxes)
original_image = draw_boxes(original_image, boxes, probs,
self.labels)
return original_image
def face_detect():
option = 'image'
net_h, net_w = 416, 416 # a multiple of 32, the smaller the faster
obj_thresh, nms_thresh = 0.4, 0.5
anchors = [
55, 69, 75, 234, 133, 240, 136, 129, 142, 363, 203, 290, 228, 184, 285,
359, 341, 260
]
def load_graph(frozen_graph_filename):
# We load the protobuf file from the disk and parse it to retrieve the
# unserialized graph_def
with open(frozen_graph_filename, "rb") as f:
graph_def = tf.GraphDef()
graph_def.ParseFromString(f.read())
# Then, we can use again a convenient built-in function to import a graph_def into the
# current default Graph
with tf.Graph().as_default() as graph:
tf.import_graph_def(graph_def,
input_map=None,
return_elements=None,
name="prefix",
op_dict=None,
producer_op_list=None)
return graph
# We use our "load_graph" function
graph = load_graph("yolo.pb")
# We can verify that we can access the list of operations in the graph
# =============================================================================
# for node in graph.as_graph_def().node:
# print node.name
# =============================================================================
# =============================================================================
# for op in graph.get_operations():
# opname=op.name
# print(opname) # <--- printing the operations snapshot below
# =============================================================================
# prefix/Placeholder/inputs_placeholder
# ...
# prefix/Accuracy/predictions
# We access the input and output nodes
x = graph.get_tensor_by_name('prefix/input_1:0')
y0 = graph.get_tensor_by_name('prefix/k2tfout_0:0')
y1 = graph.get_tensor_by_name('prefix/k2tfout_1:0')
y2 = graph.get_tensor_by_name('prefix/k2tfout_2:0')
# We launch a Session
with tf.Session(graph=graph) as sess:
if option == 'image':
img = cv2.imread('1.jpg')
cv2.resize(img, (640, 480))
img_h, img_w, _ = img.shape
# compute the predicted output for test_x
batch_input = preprocess_input(img, net_h, net_w) #416x416x3
inputs = np.zeros((1, net_h, net_w, 3), dtype='float32')
inputs[0] = batch_input
net_output = sess.run([y0, y1, y2],
feed_dict={x: inputs}) # output=1x13x13x18
batch_boxes = get_yolo_boxes(net_output, img_h, img_w, net_h,
net_w, anchors, obj_thresh,
nms_thresh)
_, _, facecen = draw_boxes(img, batch_boxes[0], ['face'],
obj_thresh)
print(facecen)
cv2.imshow('image with bboxes', img)
cv2.imwrite('jietu_out.jpg', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
elif option == 'webcam':
video_reader = cv2.VideoCapture('last.mp4')
fourcc = cv2.VideoWriter_fourcc(*'DIVX')
out = cv2.VideoWriter('last_out.mp4', fourcc, 20.0, (640, 480))
while (video_reader.isOpened()):
ret_val, img = video_reader.read()
#print(ret_val)
if ret_val == True:
img = cv2.resize(img, (640, 480))
img_h, img_w, _ = img.shape
#print ("image size is:",img_h,",",img_w)
# compute the predicted output for test_x
batch_input = preprocess_input(img, net_h,
net_w) #416x416x3
inputs = np.zeros((1, net_h, net_w, 3), dtype='float32')
inputs[0] = batch_input
net_output = sess.run([y0, y1, y2],
feed_dict={x: inputs
}) # output=1x13x13x18
batch_boxes = get_yolo_boxes(net_output, img_h, img_w,
net_h, net_w, anchors,
obj_thresh, nms_thresh)
_, _, facecen = draw_boxes(img, batch_boxes[0], ['face'],
obj_thresh)
print(facecen)
out.write(img)
cv2.imshow('image with bboxes', img)
#yield facecen
if cv2.waitKey(1) == 27:
break # esc to quit
else:
break
video_reader.release()
out.release()
cv2.destroyAllWindows()
PROJECT_ROOT = os.path.dirname(__file__) #获取当前目录
print(PROJECT_ROOT)
IMAGE_FOLDER = os.path.join(PROJECT_ROOT, "data", "test", "*.png")
img_fnames = glob.glob(IMAGE_FOLDER)
imgs = [] #存放图片
for fname in img_fnames: #读取图片
img = cv2.imread(fname)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
imgs.append(img)
yolo_v3 = Yolonet(n_classes=len(LABELS))
yolo_v3.load_weights(save_fname + ".h5") #将训练好的模型载入
import numpy as np
i = 0
for img in imgs: #依次传入模型
boxes, labels, probs = yolo_v3.detect(img, COCO_ANCHORS, imgsize)
print(boxes, labels, probs)
image = draw_boxes(img,
boxes,
labels,
probs,
class_labels=LABELS,
desired_size=400)
image = np.asarray(image, dtype=np.uint8)
plt.imsave("{}.jpg".format(i), image)
i += 1
plt.imshow(image)
plt.show()