Image Recogniser (Low-level API Tutorial)

SqueezeNet 1000 Classes Image Classifier

This chapter introduces how to build an image classifier by using low-level API only.

  • Tested on Ubuntu 16.04

Path and files

  • Python：SungemSDK-Python/examples/apps/ImageRecognition/ImageRecognition.py
  • Model file：SungemSDK-Python/examples/graphs/graph_sz

Image Recognition

run image recogniser with the following command under ImageRecognition directory

~/SungemSDK/examples/apps/ImageRecognition$ python3 ImageRecognition.py

The region-of-interest is highlighted with a cyan bounding box. You can adjust the size by pressing ‘w’ and ’s’.

Get Top5 classification result, as shown in the example. Image Recognition

No message displayed on the console as we are using low-level API

Parameter setting

Initialisation

# Load device
devices = hs.EnumerateDevices()
dev = hs.Device(devices[0])
dev.OpenDevice()

# Load CNN model
with open('../../graphs/graph_sz', mode='rb') as f:
	b = f.read()
graph = dev.AllocateGraph(b)
dim = (227,227)

# Load class labels
classes=np.loadtxt('../../misc/image_category.txt',str,delimiter='\t')

# Set camera mode
if WEBCAM: video_capture = cv2.VideoCapture(0)
Image classification

# Crop ROI
sz = image_raw.shape
cx = int(sz[0]/2)
cy = int(sz[1]/2)
ROI = int(sz[0]*ROI_ratio)
cropped = image_raw[cx-ROI:cx+ROI,cy-ROI:cy+ROI,:]

# Preprocessing
cropped = cropped.astype(np.float32)
cropped[:,:,0] = (cropped[:,:,0] - 104)
cropped[:,:,1] = (cropped[:,:,1] - 117)
cropped[:,:,2] = (cropped[:,:,2] - 123)

# Load to HS device and sort result
graph.LoadTensor(cv2.resize(cropped,dim).astype(np.float16), 'user object')
output, userobj = graph.GetResult()
output_label = output.argsort()[::-1][:5]
Visualisation

# Display outcome
for i in range(5):
	label = re.search("n[0-9]+\s([^,]+)", classes[output_label[i]]).groups(1)[0]
	cv2.putText(image_raw, "%s %0.2f %%" % (label, output[output_label[i]]*100), (20, 50+i*30), cv2.FONT_HERSHEY_SIMPLEX, 0.7, (0, 255, 255), 2)

# Plot bounding box
cv2.rectangle(image_raw, (cy-ROI, cx-ROI), (cy+ROI, cx+ROI),(255,255,0), 5)

# Display result image
cv2.imshow('Result',image_raw)

# Detect keyboard press
key = cv2.waitKey(1)
if key == ord('w'):
	ROI_ratio += 0.1
elif key == ord('s'):
	ROI_ratio -= 0.1
if ROI_ratio < 0.1:
	ROI_ratio = 0.1