🗼 Industrial Tower Defect Detection & Localization Model
100% Custom Mathematical Machine Learning Architecture (Pure PyTorch & NumPy)
This model is an industrial-grade Computer Vision detector for power transmission and telecommunication tower inspection, engineered from mathematical first principles in pure PyTorch with zero third-party YOLO/Ultralytics dependencies.
🎯 Target Defect Classes
| Class ID | Defect Name | Severity | Description |
|---|---|---|---|
0 |
rust |
HIGH |
Surface oxidation, flaking paint, and steel corrosion |
1 |
missing_bolt |
CRITICAL |
Unbolted joint holes and loosened fasteners on gusset plates |
2 |
veg |
MEDIUM |
Climbing vegetation, weeds, and tree branches reaching tower legs |
3 |
bird_nest |
MEDIUM |
Avian twig nests on antenna brackets and crossarms |
4 |
damaged_insulator |
CRITICAL |
Chipped, broken porcelain glass sheds or flashover burn marks |
🎨 Rust-Colored Tower Paint & Color-Invariance Defense
Towers painted with red oxide primer, reddish-brown anti-corrosion paint, or Cor-Ten weathering steel often fool traditional models into producing false rust detections.
This model employs texture-centric feature discrimination:
- Color-Invariance Training: Trained with random desaturation (grayscale) and hue jitter to prevent relying on red/brown color channels alone.
- Morphology Over Hue: Convolutions prioritize surface pitting, peeling blisters, and edge irregularities over uniform paint sheen.
- Hard Negative Background Training: Unlabeled clean painted metal surfaces train objectness logits to zero ($y_{\text{obj}} = 0$).
🧠 Neural Architecture
- Backbone: Deep Residual Convolutional Network with Squeeze-and-Excitation (SE) Channel Attention, SPPF (Spatial Pyramid Pooling - Fast), and 4-stage multi-scale downsampling ($P_3$ Stride 8, $P_4$ Stride 16, $P_5$ Stride 32) with SiLU activations.
- Neck: 3-Scale Feature Pyramid Network & Path Aggregation Network (FPN + PAN) with dual-pathway lateral fusion across $P_3, P_4, P_5$.
- Head: Triple decoupled prediction heads separating bounding box regression $[t_x, t_y, t_w, t_h]$, objectness $t_{\text{obj}}$, and class probabilities.
- Loss: Multi-task Complete IoU (CIoU) regression + $\alpha$-Balanced Focal Loss for dense background suppression + Scale-aware target routing + Multi-class Cross-Entropy (CE).
🚀 Quick Usage in Python
1. Load with Pure PyTorch
import torch
from model import RawCustomDetector, decode_predictions, custom_nms
from PIL import Image
import numpy as np
# Load architecture & weights
model = RawCustomDetector(num_classes=5)
checkpoint = torch.load("best.pt", map_location="cpu")
model.load_state_dict(checkpoint["model_state_dict"] if "model_state_dict" in checkpoint else checkpoint)
model.eval()
# Preprocess image
img = Image.open("tower_photo.jpg").convert("RGB").resize((640, 640))
tensor = torch.from_numpy(np.array(img, dtype=np.float32) / 255.0).permute(2, 0, 1).unsqueeze(0)
# Forward pass & Decode
with torch.no_grad():
preds = model(tensor)
boxes, scores, classes = decode_predictions(preds, conf_thresh=0.25)[0]
detections = custom_nms(boxes, scores, classes, iou_thresh=0.45)
print("Detected defects:", detections)
2. Load with ONNX Runtime (Cross-Platform)
import onnxruntime as ort
import numpy as np
session = ort.InferenceSession("best_640x640.onnx")
dummy_input = np.random.randn(1, 3, 640, 640).astype(np.float32)
outputs = session.run(None, {"images": dummy_input})
print("P3 Stride 8 shape (Small Defects - Bolts):", outputs[0].shape)
print("P4 Stride 16 shape (Medium Defects - Nests/Insulators):", outputs[1].shape)
print("P5 Stride 32 shape (Large Defects - Rust/Veg):", outputs[2].shape)
- Downloads last month
- 33