Computer Vision Preprocessing: The Missing 20% Before Model Training

Computer Vision Models Fail Quietly Before Training

People blame model architecture for poor vision results, but a large share of failures starts in preprocessing inconsistencies: crop policy drift, color normalization mismatch, and label noise.

Step 1: Normalize images through one deterministic pipeline

import cv2
import numpy as np

def preprocess(img, size=(224, 224)):
    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
    img = cv2.resize(img, size, interpolation=cv2.INTER_AREA)
    img = img.astype(np.float32) / 255.0
    return img

Step 2: Make label validation explicit

def validate_labels(rows):
    allowed = {"cat", "dog", "bird"}
    bad = [r for r in rows if r["label"] not in allowed]
    if bad:
        raise ValueError(f"invalid labels: {len(bad)}")

Step 3: Track dataset fingerprints

find data/train -type f -print0 | sort -z | xargs -0 shasum > data/train.sha1

Pitfalls

• Different resize policies between training and inference.
• Manual label edits with no audit trail.
• No dataset fingerprinting before retraining.

Verification

• Preprocess outputs are identical across repeated runs.
• Invalid labels fail pipeline before training starts.
• Dataset hash changes are visible in experiment logs.