Intelligent Processing Technology for Computer Image Recognition Based on Machine Vision

Abstract

Addressing the issues of data scarcity and material illumination drift in current industrial quality inspection, this paper studies a “physical-visual-semantic” (PVS) trimodal collaborative framework. Based on texture-aware mask autoencoders, high-difficulty sample alignment, and low-sample transfer methods, low-sample initialization and small-sample transfer are achieved. A Vision Transformer-Large (ViT-L) and Convolutional Neural Network (CNN) parallel encoder is constructed, and texture complexity is calculated using Sobel gradients and local entropy. Then, based on the dynamic mask probability of Sigmoid, only the masked labels are reconstructed. Based on the physical losses of L2 pixels (L2) and L1 (L1), and sparse representation based on the cosine consistency criterion, low-sample initialization and small-sample transfer under unlabeled conditions are achieved. Subsequently, the visual average pooling vector, BERT defect cue vector, and physical statistical vector are fed into the alignment trimodal framework with a hard-negative sample penalty. In the subset zero-shot test of MVTec-AD, the proposed method achieved Top-1, mAP@0.5, and microcrack recall rates of 82.8%, 78.8%, and 94.4%, respectively. Under 5-shot conditions, the accuracy reached 86.3%, and the cross-material Area Under the ROC Curve (AUC-ROC) reached 89.2%, meeting the real-time requirements of the production line.