Auditory brainstem response (ABR) testing is widely used in hearing screening and clinical diagnostics. In routine practice, ABR recordings are often interpreted from graphical reports rather than raw digital signals, making wave V identification and hearing-threshold estimation challenging, particularly near threshold. This study presents a computer-vision framework that estimates hearing thresholds and wave V latencies directly from ABR report images without reconstructing one-dimensional signals. YOLO and RF-DETR object detectors were evaluated in combination with physiologically constrained post-processing that selects consistent detections across stimulus intensity levels. The best-performing configuration, based on an optimized YOLO26M detector, achieved 92.76% accuracy in identifying wave V presence, a mean absolute wave V latency error of 0.28 ms, and 94.3% agreement with expert-defined hearing thresholds within ±10 dB. Post-processing improved threshold agreement by enforcing consistency across intensity levels. These findings demonstrate the feasibility of extracting clinically relevant information directly from graphical ABR reports and suggest potential utility as decision support for audiologists. Further evaluation across clinical centers, recording devices, and report formats is required to establish generalizability.