Tamper-Evident Industrial Visual Inspection Using Blockchain and Feature-Level Hashing

Images captured by automated optical inspection stations are the evidence behind accept and reject decisions in manufacturing, yet they are stored as ordinary files that can be replaced or retouched after the fact. A cryptographic digest exposes any change, but it cannot distinguish a forged defect region from the routine brightness adjustment or recompression that an image undergoes when it is archived, viewed or exported, so in practice it raises alarms on nearly every legitimate copy. This paper presents a tamper-evident inspection record that keeps an exact digest for strict audits and adds a block-level feature fingerprint for tolerant verification. Each image is divided into 64 blocks, and every block is summarised by normalised intensity, gradient energy, edge density and a contrast-thresholded local binary pattern histogram. The quantised fingerprint and the digest form a record whose hash is committed to an append-only hash chain through Merkle batching. At audit time a trimmed regression removes globally uniform feature changes, each feature group is scaled by its own benign variability, and blocks whose fused score exceeds a calibrated threshold are reported as tampered. In a simulation with 200 synthetic steel-surface images, 2400 benign images and 2342 forgeries, the scheme detected 98.2 percent of defect insertions, removals and substitutions at a 0.67 percent benign false-alarm rate, against 34.8 percent for a global perceptual hash. Tampered blocks were located with 94.1 percent precision. The method is not robust to small translations or blurring, which caused false alarms on most images and define the main direction for future work.