OpenVINO提供的場景文字檢測模型準確率是非常的高，完全可以達到實用級別，其實OpenVINO還提供了另外一個場景文字識別的模型，總體使用下來的感覺是沒有場景文字檢測那么靠譜，而且只支持英文字母與數字識別，不支持中文，不得不說是一個小小遺憾，但是對比較干凈的檔案影像，它的識別準確率還是相當的高，速度也比較快，基本上都在毫秒基本出結果。
模型介紹
文本識別(OCR)模型采用的網路架構為基礎網路+雙向LSTM，其中基礎網路選擇的是VGG16，字母識別是非大小寫敏感的，26個字母+10個數字總計36個字符。其網路結構類似如下：
模型輸入結構為：
[BxCxHxW]=1x1x32x120
其中B表示批次、C表示通道、H表示高度、W表示寬度
模型輸出結果為：
[WxBxL] = 30x1x37
其中B表示批次、W表示輸出序列長度、L表示各個37個字符各自得分，其中第37個是#
輸出部分的決議基于CTC貪心解碼方式。
代碼實作
加載模型
# 加載IR
log.info("Reading IR...")
net = IENetwork(model=model_xml, weights=model_bin)
text_net = IENetwork(model=text_xml, weights=text_bin)
場景文字檢測
# image = cv2.imread("D:/images/openvino_ocr.png");
image = cv2.imread("D:/images/cover_01.jpg");
cv2.imshow("image", image)
inf_start = time.time()
in_frame = cv2.resize(image, (w, h))
in_frame = in_frame.transpose((2, 0, 1)) # Change data layout from HWC to CHW
in_frame = in_frame.reshape((n, c, h, w))
exec_net.infer(inputs={input_blob: in_frame})ROI截取與文字識別
x, y, width, height = cv2.boundingRect(contours[c])
roi = image[y-5:y+height+10,x-5:x+width+10,:]
gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
text_roi = cv2.resize(gray, (tw, th))
text_roi = np.expand_dims(text_roi, 2)
text_roi = text_roi.transpose((2, 0, 1))
text_roi = text_roi.reshape((tn, tc, th, tw))
text_exec_net.infer(inputs={input_blob: text_roi})
text_out = text_exec_net.requests[0].outputs[text_out_blob]
CTC決議結果
# 決議輸出text
ocrstr = ""
prev_pad = False;
for i in range(text_out.shape[0]):
ctc = text_out[i]
ctc = np.squeeze(ctc, 0)
index, prob = ctc_soft_max(ctc)
if alphabet[index] == '#':
prev_pad = True
else:
if len(ocrstr) == 0 or prev_pad or (len(ocrstr) > 0 and alphabet[index] != ocrstr[-1]):
prev_pad = False
ocrstr += alphabet[index]
輸出文字檢測與識別結果
# 顯示識別結果
print("result: %s"%ocrstr)
cv2.drawContours(image, [box], 0, (0, 255, 0), 2)
cv2.putText(image, ocrstr, (x, y), cv2.FONT_HERSHEY_COMPLEX, 0.75, (255, 0, 0), 1)
最后送上整個演示代碼
def demo():
# 加載MKLDNN - CPU Target
log.basicConfig(format="[ %(levelname)s ] %(message)s", level=log.INFO, stream=sys.stdout)
plugin = IEPlugin(device="CPU", plugin_dirs=plugin_dir)
plugin.add_cpu_extension(cpu_extension)
# 加載IR
log.info("Reading IR...")
net = IENetwork(model=model_xml, weights=model_bin)
text_net = IENetwork(model=text_xml, weights=text_bin)
if plugin.device == "CPU":
supported_layers = plugin.get_supported_layers(net)
not_supported_layers = [l for l in net.layers.keys() if l not in supported_layers]
if len(not_supported_layers) != 0:
log.error("Following layers are not supported by the plugin for specified device {}:\n {}".
format(plugin.device, ', '.join(not_supported_layers)))
log.error("Please try to specify cpu extensions library path in demo's command line parameters using -l "
"or --cpu_extension command line argument")
sys.exit(1)
# 獲取輸入輸出層
input_blob = next(iter(net.inputs))
outputs = iter(net.outputs)
# 獲取多個輸出層名稱
out_blob = next(outputs)
second_blob = next(outputs)
log.info("Loading IR to the plugin...")
print("pixel output: %s, link output: %s \n"%(out_blob, second_blob))
text_input_blob = next(iter(text_net.inputs))
text_out_blob = next(iter(text_net.outputs))
print("text_out_blob : %s"%text_out_blob)
# 創建可執行網路
exec_net = plugin.load(network=net)
text_exec_net = plugin.load(network=text_net)
# Read and pre-process input image
n, c, h, w = net.inputs[input_blob].shape
tn, tc, th, tw = text_net.inputs[text_input_blob].shape
del net
del text_net
log.info("Starting inference in async mode...")
log.info("To switch between sync and async modes press Tab button")
log.info("To stop the demo execution press Esc button")
image = cv2.imread("D:/images/openvino_ocr.png");
# image = cv2.imread("D:/images/cover_01.jpg");
cv2.imshow("image", image)
inf_start = time.time()
in_frame = cv2.resize(image, (w, h))
in_frame = in_frame.transpose((2, 0, 1)) # Change data layout from HWC to CHW
in_frame = in_frame.reshape((n, c, h, w))
exec_net.infer(inputs={input_blob: in_frame})
inf_end = time.time()
det_time = inf_end - inf_start
# 獲取輸出
res1 = exec_net.requests[0].outputs[out_blob]
res2 = exec_net.requests[0].outputs[second_blob]
# 降維
res1 = np.squeeze(res1, 0)
res2 = np.squeeze(res2, 0)
# 矩陣轉置
res1 = res1.transpose((1, 2, 0))
res2 = res2.transpose((1, 2, 0))
h, w = res1.shape[:2]
print(res1.shape)
print(res2.shape)
# 文本與非文本像素
pixel_mask = np.zeros((h, w), dtype=np.uint8)
# 決議輸出結果
res1 = soft_max(res1)
# 像素分割
for row in range(h):
for col in range(w):
pv2 = res1[row, col, 1]
if pv2 > 0.50:
pixel_mask[row, col] = 255
se = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 1))
mask = cv2.morphologyEx(pixel_mask, cv2.MORPH_CLOSE, se)
cv2.imshow("text mask", mask)
cv2.imwrite("D:/mask.png", mask)
# 后處理，檢測框
h, w = image.shape[:2]
mask = cv2.resize(mask, (w, h))
contours, hierarchy = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
for c in range(len(contours)):
rect = cv2.minAreaRect(contours[c])
box = cv2.boxPoints(rect)
box = np.int0(box)
x, y, width, height = cv2.boundingRect(contours[c])
roi = image[y-5:y+height+10,x-5:x+width+10,:]
gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
text_roi = cv2.resize(gray, (tw, th))
text_roi = np.expand_dims(text_roi, 2)
text_roi = text_roi.transpose((2, 0, 1))
text_roi = text_roi.reshape((tn, tc, th, tw))
text_exec_net.infer(inputs={input_blob: text_roi})
text_out = text_exec_net.requests[0].outputs[text_out_blob]
# 決議輸出text
ocrstr = ""
prev_pad = False;
for i in range(text_out.shape[0]):
ctc = text_out[i]
ctc = np.squeeze(ctc, 0)
index, prob = ctc_soft_max(ctc)
if alphabet[index] == '#':
prev_pad = True
else:
if len(ocrstr) == 0 or prev_pad or (len(ocrstr) > 0 and alphabet[index] != ocrstr[-1]):
prev_pad = False
ocrstr += alphabet[index]
# 顯示識別結果
print("result: %s"%ocrstr)
cv2.drawContours(image, [box], 0, (0, 255, 0), 2)
cv2.putText(image, ocrstr, (x, y), cv2.FONT_HERSHEY_COMPLEX, 0.75, (255, 0, 0), 1)
inf_time_message = "Inference time: {:.3f} ms， FPS:{:.3f}".format(det_time * 1000, 1000 / (det_time * 1000))
cv2.putText(image, inf_time_message, (15, 15), cv2.FONT_HERSHEY_COMPLEX, 0.5, (255, 255, 0), 1)
cv2.imshow("result", image)
cv2.imwrite("D:/result.png", image)
cv2.waitKey(0)
# 釋放資源
cv2.destroyAllWindows()
del exec_net
del plugin
演示效果
OCR識別輸出 - 效果一
OCR識別輸出 - 效果二
總結：
發現對特定的應用場景，特別是一些檔案化的影像，這個模型識別還比較準確，對很多其它的應用場景，比如身份證、各種卡號識別，發現誤識別率很高，現如這些場景需要專項訓練的模型！
轉載請註明出處，本文鏈接：https://www.uj5u.com/yidong/175299.html
標籤：英特爾技術