Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/oktopuzSlid/detectorPlacas/llms.txt

Use this file to discover all available pages before exploring further.

DetectorPlacas uses SSD (Single Shot MultiBox Detector) with MobileNet v1 as the feature extractor, implemented through TensorFlow 1.x’s frozen graph inference API. Rather than running a two-stage detector like Faster R-CNN, SSD predicts bounding boxes and class scores for all anchor locations in a single forward pass — making it fast enough for real-time webcam use on modest hardware. This page covers the model’s input and output shapes, the tensor names the scripts bind to, how the frozen graph is loaded into a TF session, and how inference results are visualized.

​
Model Overview

PropertyValue
ArchitectureSSD — Single Shot MultiBox Detector
Feature extractorMobileNet v1 (depth_multiplier=1.0, min_depth=16)
Input resolution300 × 300 RGB (fixed shape resizer)
Number of classes1 — license plate only
ActivationRELU_6 throughout all conv layers
Batch normalizationApplied to every convolutional layer (decay=0.9997, ε=0.001)
Anchor layers6 SSD feature map layers
Anchor scale range0.2 – 0.95 of input size
Anchor aspect ratios5 per location: 1.0, 2.0, 0.5, 3.0, 0.3333
Box predictorConvolutional, kernel_size=1, no dropout at inference
Score conversionSIGMOID
Post-processingNon-Maximum Suppression, IoU threshold 0.6, up to 100 detections

​
Loading the Frozen Graph

All three detection scripts use identical code to load the model from frozen_inference_graph.pb. The protobuf file is read from disk, parsed into a GraphDef, imported into a fresh tf.Graph(), and a tf.Session() is created bound to that graph. 
MODEL_NAME = 'inference_graph'
CWD_PATH = os.getcwd()
PATH_TO_CKPT = os.path.join(CWD_PATH, MODEL_NAME, 'frozen_inference_graph.pb')

detection_graph = tf.Graph()
with detection_graph.as_default():
    od_graph_def = tf.GraphDef()
    with tf.gfile.GFile(PATH_TO_CKPT, 'rb') as fid:
        serialized_graph = fid.read()
        od_graph_def.ParseFromString(serialized_graph)
        tf.import_graph_def(od_graph_def, name='')

    sess = tf.Session(graph=detection_graph)
The name='' argument to tf.import_graph_def() preserves the original tensor names (e.g., image_tensor:0) without adding a namespace prefix. The resulting sess object is kept open for the lifetime of the script and reused for every inference call.
This is TensorFlow 1.x session-based inference. TF 2.x removed both tf.Session() and tf.GraphDef(). If you need to migrate, the saved_model/saved_model.pb export bundled with the project can be loaded in TF 2.x using tf.saved_model.load(), which returns a callable with the same input/output semantics.

​
Input Preparation

OpenCV loads images in BGR channel order by default. The model expects RGB input, so every script converts the color channel order before expanding the batch dimension. 
image_rgb = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image_expanded = np.expand_dims(image_rgb, axis=0)
# Shape: [1, H, W, 3] — batch dimension required by the model
np.expand_dims(..., axis=0) inserts the batch dimension at position 0, turning a [H, W, 3] array into a [1, H, W, 3] array. The model’s fixed-shape resizer handles the resize to 300 × 300 internally, so the input does not need to be pre-resized.

​
Inference Tensors

All five tensors are retrieved from the loaded graph by name immediately after the session is created. The same code appears in all three detection scripts.
image_tensor      = detection_graph.get_tensor_by_name('image_tensor:0')
detection_boxes   = detection_graph.get_tensor_by_name('detection_boxes:0')
detection_scores  = detection_graph.get_tensor_by_name('detection_scores:0')
detection_classes = detection_graph.get_tensor_by_name('detection_classes:0')
num_detections    = detection_graph.get_tensor_by_name('num_detections:0')
Tensor nameDirectionShapeNotes
image_tensor:0Input[batch, height, width, 3]RGB uint8 pixel values. The model internally resizes to 300 × 300.
detection_boxes:0Output[batch, num_detections, 4]Normalized coordinates [ymin, xmin, ymax, xmax] in range [0, 1].
detection_scores:0Output[batch, num_detections]Confidence score per detection box, as a float in [0, 1].
detection_classes:0Output[batch, num_detections]Class index per detection. 1 = license plate. Cast to int32 with .astype(np.int32).
num_detections:0Output[batch]Number of valid detections returned for each image in the batch.

​
Running Inference

The sess.run() call evaluates all four output tensors in a single forward pass, feeding the prepared image batch via feed_dict. 
(boxes, scores, classes, num) = sess.run(
    [detection_boxes, detection_scores, detection_classes, num_detections],
    feed_dict={image_tensor: image_expanded})
All four output arrays are returned simultaneously. Each has a leading batch dimension of size 1 (since image_expanded contains only one image), which must be removed before passing to the visualization utility.

​
Visualization

After inference, vis_util.visualize_boxes_and_labels_on_image_array() draws bounding boxes, class labels, and confidence scores directly onto the original image array in place. 
vis_util.visualize_boxes_and_labels_on_image_array(
    image,
    np.squeeze(boxes),
    np.squeeze(classes).astype(np.int32),
    np.squeeze(scores),
    category_index,
    use_normalized_coordinates=True,
    line_thickness=8,
    min_score_thresh=0.65)  # 0.65 for image, 0.55 for video, 0.50 for webcam
ArgumentMeaning
np.squeeze(boxes)Removes the batch dimension, yielding shape [num_detections, 4]
np.squeeze(classes).astype(np.int32)Removes the batch dimension and casts class indices to int32 for category lookup
np.squeeze(scores)Removes the batch dimension, yielding shape [num_detections]
use_normalized_coordinates=TrueTells vis_util that box coordinates are in [0, 1] and should be scaled to pixel space by the function
line_thicknessBorder width in pixels for the drawn bounding boxes
min_score_threshDetections with a confidence score below this threshold are not drawn. The threshold is deliberately loosest for webcam (0.50) to reduce misses in real-time use, and tightest for static images (0.65) where precision matters more.

​
Model Artifacts

The project directory contains the following model files. The primary inference artifact is the frozen protobuf graph; the checkpoint files are retained for fine-tuning or export.
FileFormatPurpose
frozen_inference_graph.pbFrozen protobuf graphPrimary inference artifact — all weights are embedded; used by all three detection scripts
model.ckpt.data-00000-of-00001TF checkpoint dataVariable values for the trained model
model.ckpt.indexTF checkpoint indexMaps variable names to their offsets in the data file
model.ckpt.metaTF checkpoint metaSerialized MetaGraphDef — graph structure and saver information
saved_model/saved_model.pbSavedModel formatTF 2.x-compatible export; loadable via tf.saved_model.load()
checkpointPlain text pointerContains model_checkpoint_path: "model.ckpt" and all_model_checkpoint_paths: "model.ckpt" — tells TF which checkpoint is current

Build docs developers (and LLMs) love

Get started for freeTalk to us