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 is a Python computer vision project that uses TensorFlow 1.x and a frozen SSD MobileNet v1 inference graph to locate and highlight vehicle license plates. Built on top of the TensorFlow Object Detection API and OpenCV, it supports three distinct detection modes — static images, pre-recorded video files, and live webcam feeds — making it practical for both batch image analysis and continuous real-time surveillance. The model was trained on a single-class dataset (license plates) and produces bounding boxes annotated with confidence scores on every frame it processes.

​
What DetectorPlacas Does

DetectorPlacas exposes three detection modes, each implemented as a self-contained Python script and each tuned with its own minimum confidence threshold to balance precision against recall for that input type.
ModeScriptConfidence ThresholdWindow Title
ImageObject_detection_image.py0.65Detector de Placas PRUEBA-IMAGEN
VideoObject_detection_video.py0.55Detector de placas- PRUEBA VIDEO
WebcamObject_detection_webcam.py0.50Detector de placas PRUEBA-WEBCAM
Image mode scans a directory for .jpg files using os.listdir() combined with endswith('.jpg'), and runs inference on each file sequentially, displaying results in an OpenCV window. The directory is set via the --ruta flag; if omitted, the current working directory is used. Video mode opens a file named prueba.mp4 from the working directory and processes it frame by frame. A slightly lower threshold of 0.55 is used to account for motion blur and varying lighting across frames. Webcam mode captures a continuous live feed from camera index 0 at a resolution of 1280×720 pixels and runs inference on every frame. The lowest threshold of 0.50 is applied to ensure plates are caught even under challenging real-time conditions. In all three modes, detections are visualised using vis_util.visualize_boxes_and_labels_on_image_array, which draws labelled bounding boxes with a line thickness of 8 px (image and webcam) or 5 px (video) directly onto the output frame.

​
Key Components

frozen_inference_graph.pb

The pre-trained SSD MobileNet v1 model exported as a frozen TensorFlow graph. Stored inside inference_graph/, this .pb file contains all weights and the full computation graph needed for inference — no separate training step is required at runtime.

labelmap.pbtxt

The label map that maps the single integer class index to its human-readable name. DetectorPlacas recognises exactly one class — license_plate — so the file contains a single item block with id: 1.

Detection Scripts

Three standalone Python scripts — Object_detection_image.py, Object_detection_video.py, and Object_detection_webcam.py — each loading the same frozen graph and label map but reading input from a different source and applying a mode-appropriate confidence threshold.

Graphical Interface

interfaz.py is a Tkinter GUI with window title "Proyecto" that presents radio buttons for IMAGEN, VIDEO, and CAMARA modes alongside a text field for the image directory path. Clicking PRUEBA dispatches the correct script via os.system, making it easy to switch modes without touching the command line.

​
Architecture Overview

Every detection script follows the same five-stage pipeline:
  1. Load the label map — label_map_util.load_labelmap reads labelmap.pbtxt, label_map_util.convert_label_map_to_categories converts it to a list of category dicts, and label_map_util.create_category_index builds the integer-to-name mapping used by the visualiser.
  2. Deserialise the frozen graph — A tf.GraphDef() object is populated by reading frozen_inference_graph.pb with tf.gfile.GFile, then imported into a dedicated tf.Graph() via tf.import_graph_def.
  3. Open a TensorFlow session — tf.Session(graph=detection_graph) creates the runtime context bound to the frozen graph.
  4. Feed the image tensor — Each frame is converted from BGR to RGB with OpenCV, expanded to a batch dimension of 1, and passed to the session via feed_dict={image_tensor: frame_expanded}.
  5. Collect outputs and visualise — The session returns four tensors:
    • detection_boxes:0 — normalised [ymin, xmin, ymax, xmax] coordinates for each detected object
    • detection_scores:0 — per-box confidence scores in [0, 1]
    • detection_classes:0 — integer class indices
    • num_detections:0 — total number of detections in the batch
    vis_util.visualize_boxes_and_labels_on_image_array overlays boxes and scores on the original frame, which is then shown with cv2.imshow.

​
Dependencies

DetectorPlacas depends on the following libraries. Ensure all are installed before running any detection script:
  • TensorFlow 1.x — the scripts use tf.app.flags, tf.Graph(), tf.Session(), and tf.GraphDef(), which are TF 1.x APIs removed or substantially changed in TF 2.x
  • OpenCV (cv2) — image and video I/O, colour-space conversion, and display
  • NumPy — tensor reshaping (np.expand_dims, np.squeeze)
  • TensorFlow Object Detection API — provides label_map_util and visualization_utils from the tensorflow/models research repository
  • Tkinter — standard-library GUI toolkit used by interfaz.py (included with most Python distributions)

​
Project Structure

All scripts resolve paths relative to the current working directory, so the following layout must be preserved exactly when running any detection script: 
project-root/
├── inference_graph/
│   ├── frozen_inference_graph.pb
│   ├── model.ckpt.data-00000-of-00001
│   ├── model.ckpt.index
│   └── model.ckpt.meta
├── labelmap.pbtxt
├── Object_detection_image.py
├── Object_detection_video.py
├── Object_detection_webcam.py
└── interfaz.py
The model training configuration (pipeline.config) specifies num_examples: 1100 for evaluation and uses coco_detection_metrics as the metrics set. These COCO detection metrics — including mean Average Precision (mAP) at various IoU thresholds — are the standard benchmarks used to measure license-plate detection accuracy for this model.

Build docs developers (and LLMs) love

Get started for freeTalk to us