DACA-Net: Detail-aware network with contrast attention for locating liquid crystal display defects

Abstract

Listen

Translate article

Liquid crystal display (LCD) defects are characterized by their tiny size and low contrast, making it difficult to accurately categorize and quantify these defects during inspection. To overcome these challenges, we proposed a detail-aware network with contrast attention (DACA-Net). First, we constructed a low-level semantic refinement (LSR) module, which restructures the cross stage partial (CSP) structure of the backbone network, enhancing the detail-aware capability for tiny objects. Second, we designed a low-level semantic deep fusion (LSDF) module that integrates detailed semantic information during feature fusion, effectively mitigating the loss of tiny objects information post deep convolution. Third, we proposed a dual-focus contrast enhancement attention (DFCEA) module, which alleviates the insufficient spatial semantic information of low-contrast objects, thereby enhancing the detection of defects under low-contrast background. Moreover, we present a representative LCD light Defect (LCDLD) dataset to address the lack of dataset in this field. We validated our model using the LCDLD and PKU-Market-Phone datasets and tested its generalizability with the PKU-Market-PCB dataset. DACA-Net achieved mean average precision (mAP) scores of 96.7%, 98.5%, and 97.1% on these datasets, respectively, surpassing 16 state-of-the-art object and defect detection models.