License Plate Detection using YOLOv4
The primary objective of this project is to implement a YOLOv4 neural network to detect and identify license plates in vehicle images. The implementation involves loading the pre-trained YOLOv4 model, processing the images, and visualizing the detection results.
Dataset
The dataset consists of images of Moroccan vehicle license plates. Each image is accompanied by a file that specifies the position of the plate in YOLO format. The images were labeled using LabelImg, a graphical image annotation tool.
YOLO Model
YOLO (You Only Look Once) is a state-of-the-art, real-time object detection system. YOLOv4, the model used in this project, is an improvement over previous versions and offers enhanced accuracy and speed.
YOLO Model Comparison
| Model Version | Description | Speed (FPS) | mAP (Mean Average Precision) |
|---|---|---|---|
| YOLOv1 | Initial version with basic features | ~45 FPS | ~63.4% |
| YOLOv2 | Improved accuracy and speed, multi-scale training | ~40 FPS | ~76.8% |
| YOLOv3 | Further improvements in accuracy, deeper architecture | ~30 FPS | ~80.0% |
| YOLOv4 | Latest version with advanced features like CSPDarknet53 | ~50 FPS | ~89.7% |
Code Explanation
1. Importing Libraries
import numpy as np
import cv2
import matplotlib.pyplot as plt
import time
This cell imports necessary libraries: numpy for numerical operations, cv2 for image processing, matplotlib for plotting, and time for measuring execution time.
2. Reading Class Labels
labels = open("obj.names").read().strip().split('\n') # list of class names
print(labels)
Reads class names from the obj.names file and prints them.
Output:
['license_plate']
3. Defining Paths and Parameters
weights_path = "yolov4-obj.weights"
configuration_path = "yolov4-obj.cfg"
probability_minimum = 0.5
threshold = 0.3
Sets the paths to the YOLO weights and configuration files and defines the minimum probability and threshold for non-maximum suppression.
4. Loading YOLO Model
network = cv2.dnn.readNetFromDarknet(configuration_path, weights_path)
Loads the YOLO model with the specified configuration and weights.
5. Getting Layer Names
layers_names_all = network.getLayerNames() # list of layers names
print(layers_names_all)
Gets and prints all layer names in the YOLO model.
Output:
('conv_0', 'bn_0', 'mish_1', 'conv_1', 'bn_1', 'mish_2', 'conv_2', 'bn_2', 'mish_3', 'identity_3', 'conv_4', 'bn_4', 'mish_5', 'conv_5', 'bn_5', 'mish_6', 'conv_6', 'bn_6', 'mish_7', 'shortcut_7', 'conv_8', 'bn_8', 'mish_9', 'concat_9', 'conv_10', 'bn_10', 'mish_11', 'conv_11', 'bn_11', 'mish_12', 'conv_12', 'bn_12', 'mish_13', 'identity_13', 'conv_14', 'bn_14', 'mish_15', 'conv_15', 'bn_15', 'mish_16', 'conv_16', 'bn_16', 'mish_17', 'shortcut_17', 'conv_18', 'bn_18', 'mish_19', 'conv_19', 'bn_19', 'mish_20', 'shortcut_20', 'conv_21', 'bn_21', 'mish_22', 'concat_22', 'conv_23', 'bn_23', 'mish_24', 'conv_24', 'bn_24', 'mish_25', 'conv_25', 'bn_25', 'mish_26', 'identity_26', 'conv_27', 'bn_27', 'mish_28', 'conv_28', 'bn_28', 'mish_29', 'conv_29', 'bn_29', 'mish_30', 'shortcut_30', 'conv_31', 'bn_31', 'mish_32', 'conv_32', 'bn_32', 'mish_33', 'shortcut_33', 'conv_34', 'bn_34', 'mish_35', 'conv_35', 'bn_35', 'mish_36', 'shortcut_36', 'conv_37', 'bn_37', 'mish_38', 'conv_38', 'bn_38', 'mish_39', 'shortcut_39', 'conv_40', 'bn_40', 'mish_41', 'conv_41', 'bn_41', 'mish_42', 'shortcut_42', 'conv_43', 'bn_43', 'mish_44', 'conv_44', 'bn_44', 'mish_45', 'shortcut_45', 'conv_46', 'bn_46', 'mish_47', 'conv_47', 'bn_47', 'mish_48', 'shortcut_48', 'conv_49', 'bn_49', 'mish_50', 'conv_50', 'bn_50', 'mish_51', 'shortcut_51', 'conv_52', 'bn_52', 'mish_53', 'concat_53', 'conv_54', 'bn_54', 'mish_55', 'conv_55', 'bn_55', 'mish_56', 'conv_56', 'bn_56', 'mish_57', 'identity_57', 'conv_58', 'bn_58', 'mish_59', 'conv_59', 'bn_59', 'mish_60', 'conv_60', 'bn_60', 'mish_61', 'shortcut_61', 'conv_62', 'bn_62', 'mish_63', 'conv_63', 'bn_63', 'mish_64', 'shortcut_64', 'conv_65', 'bn_65', 'mish_66', 'conv_66', 'bn_66', 'mish_67', 'shortcut_67', 'conv_68', 'bn_68', 'mish_69', 'conv_69', 'bn_69', 'mish_70', 'shortcut_70', 'conv_71', 'bn_71', 'mish_72', 'conv_72', 'bn_72', 'mish_73', 'shortcut_73', 'conv_74', 'bn_74', 'mish_75', 'conv_75', 'bn_75', 'mish_76', 'shortcut_76', 'conv_77', 'bn_77', 'mish_78', 'conv_78', 'bn_78', 'mish_79', 'shortcut_79', 'conv_80', 'bn_80', 'mish_81', 'conv_81', 'bn_81', 'mish_82', 'shortcut_82', 'conv_83', 'bn_83', 'mish_84', 'concat_84', 'conv_85', 'bn_85', 'mish_86', 'conv_86', 'bn_86', 'mish_87', 'conv_87', 'bn_87', 'mish_88', 'identity_88', 'conv_89', 'bn_89', 'mish_90', 'conv_90', 'bn_90', 'mish_91', 'conv_91', 'bn_91', 'mish_92', 'shortcut_92', 'conv_93', 'bn_93', 'mish_94', 'conv_94', 'bn_94', 'mish_95', 'shortcut_95', 'conv_96', 'bn_96', 'mish_97', 'conv_97', 'bn_97', 'mish_98', 'shortcut_98', 'conv_99', 'bn_99', 'mish_100', 'conv_100', 'bn_100', 'mish_101', 'shortcut_101', 'conv_102', 'bn_102', 'mish_103', 'concat_103', 'conv_104', 'bn_104', 'mish_105', 'conv_105', 'bn_105', 'leaky_106', 'conv_106', 'bn_106', 'leaky_107', 'conv_107', 'bn_107', 'leaky_108', 'pool_108', 'identity_109', 'pool_110', 'identity_111', 'pool_112', 'concat_113', 'conv_114', 'bn_114', 'leaky_115', 'conv_115', 'bn_115', 'leaky_116', 'conv_116', 'bn_116', 'leaky_117', 'conv_117', 'bn_117', 'leaky_118', 'upsample_118', 'identity_119', 'conv_120', 'bn_120', 'leaky_121', 'concat_121', 'conv_122', 'bn_122', 'leaky_123', 'conv_123', 'bn_123', 'leaky_124', 'conv_124', 'bn_124', 'leaky_125', 'conv_125', 'bn_125', 'leaky_126', 'conv_126', 'bn_126', 'leaky_127', 'conv_127', 'bn_127', 'leaky_128', 'upsample_128', 'identity_129', 'conv_130', 'bn_130', 'leaky_131', 'concat_131', 'conv_132', 'bn_132', 'leaky_133', 'conv_133', 'bn_133', 'leaky_134', 'conv_134', 'bn_134', 'leaky_135', 'conv_135', 'bn_135', 'leaky_136', 'conv_136', 'bn_136', 'leaky_137', 'conv_137', 'bn_137', 'leaky_138', 'conv_138', 'permute_139', 'yolo_139', 'identity_140', 'conv_141', 'bn_141', 'leaky_142', 'concat_142', 'conv_143', 'bn_143', 'leaky_144', 'conv_144', 'bn_144', 'leaky_145', 'conv_145', 'bn_145', 'leaky_146', 'conv_146', 'bn_146', 'leaky_147', 'conv_147', 'bn_147', 'leaky_148', 'conv_148', 'bn_148', 'leaky_149', 'conv_149', 'permute_150', 'yolo_150', 'identity_151', 'conv_152', 'bn_152', 'leaky_153', 'concat_153', 'conv_154', 'bn_154', 'leaky_155', 'conv_155', 'bn_155', 'leaky_156', 'conv_156', 'bn_156', 'leaky_157', 'conv_157', 'bn_157', 'leaky_158', 'conv_158', 'bn_158', 'leaky_159', 'conv_159', 'bn_159', 'leaky_160', 'conv_160', 'permute_161', 'yolo_161')
6. Output Layers
print(network.getUnconnectedOutLayers())
layers_names_output = [layers_names_all[i - 1] for i in network.getUnconnectedOutLayers()]
print(layers_names_output)
Identify the names of the output layers required by the YOLO algorithm.
Output:
[327 353 379]
['yolo_139', 'yolo_150', 'yolo_161']
7. Reading and Displaying the Image
image_input = cv2.imread("WhatsApp Image 2023-10-28 at 18.15.55_8d3d7646.jpg")
image_input_shape = image_input.shape
print(image_input_shape)
plt.imshow(cv2.cvtColor(image_input, cv2.COLOR_BGR2RGB))
plt.show()
Reads the input image and displays its shape. Then, it displays the image.
Output shape:
(1280, 1024, 3)
8. Creating Blob from Image
blob = cv2.dnn.blobFromImage(image_input, 1 / 255.0, (416, 416), swapRB=True, crop=False)
print(image_input.shape)
print(blob.shape)
Creates a blob from the image to pass to the YOLO model and prints the shapes.
Output:
(1280, 1024, 3)
(1, 3, 416, 416)
9. Displaying the Blob
blob_to_show = blob[0, :, :, :].transpose(1, 2, 0)
print(blob_to_show.shape)
plt.imshow(blob_to_show)
plt.show()
Displays the blob.
Output shape:
(416, 416, 3)
10. Forward Pass through YOLO
network.setInput(blob)
start = time.time()
output_from_network = network.forward(layers_names_output)
end = time.time()
print('YOLO v4 took {:.3f} seconds'.format(end - start))
Performs a forward pass through the YOLO network and measures the time taken.
Output:
YOLO v4 took 1.935 seconds
11. Processing YOLO Output
print(type(output_from_network))
print(type(output_from_network[0]))
Prints the types of the outputs.
Output:
<class 'tuple'>
<class 'numpy.ndarray'>
12. Generating Colors for Classes
np.random.seed(42)
colours = np.random.randint(0, 255, size=(len(labels), 3), dtype='uint8')
print(colours.shape)
print(colours[0])
Generates random colors for bounding boxes.
Output:
(1, 3)
[102 220 225]
13. Detecting Objects
bounding_boxes = []
confidences = []
class_numbers = []
h, w = image_input_shape[:2]
print(h, w)
Initializes lists for bounding boxes, confidences, and class numbers and prints the image dimensions.
Output:
1280 1024
14. Extracting Bounding Boxes and Confidences
for result in output_from_network:
for detection in result:
scores = detection[5:]
class_current = np.argmax(scores)
confidence_current = scores[class_current]
if confidence_current > probability_minimum:
box_current = detection[0:4] * np.array([w, h, w, h])
x_center, y_center, box_width, box_height = box_current.astype('int')
x_min = int(x_center - (box_width / 2))
y_min = int(y_center - (box_height / 2))
bounding_boxes.append([x_min, y_min, int(box_width), int(box_height)])
confidences.append(float(confidence_current))
class_numbers.append(class_current)
Extracts bounding boxes and confidences for detected objects.
15. Applying Non-Maximum Suppression
results = cv2.dnn.NMSBoxes(bounding_boxes, confidences, probability_minimum, threshold)
Applies Non-Maximum Suppression to filter out overlapping bounding boxes.
16. Displaying Detected Labels
for i in range(len(class_numbers)):
print(labels[int(class_numbers[i])])
with open('found_labels.txt', 'w') as f:
for i in range(len(class_numbers)):
f.write(labels[int(class_numbers[i])])
Prints and saves detected labels.
Output:
license_plate
license_plate
17. Drawing Bounding Boxes
if len(results) > 0:
for i in results.flatten():
x_min, y_min = bounding_boxes[i][0], bounding_boxes[i][1]
box_width, box_height = bounding_boxes[i][2], bounding_boxes[i][3]
colour_box_current = [int(j) for j in colours[class_numbers[i]]]
cv2.rectangle(image_input, (x_min, y_min), (x_min + box_width, y_min + box_height), colour_box_current, 2)
text_box_current = '{} : {:.2f}%'.format(labels[int(class_numbers[i])], confidences[i])
cv2.putText(image_input, text_box_current, (x_min, y_min - 7), cv2.FONT_HERSHEY_SIMPLEX, 0.5, colour_box_current, 2)
plt.imshow(cv2.cvtColor(image_input, cv2.COLOR_BGR2RGB))
plt.axis('off')
plt.show()
Draws bounding boxes and labels on the image and displays it.
