The growing deployment of photovoltaic systems has made reliable solar panel monitoring essential for sustainable operation, lower maintenance costs, and improved energy efficiency. Surface conditions such as breakage, dust accumulation, shadowing, and other visible abnormalities can reduce panel performance and shorten service life. Conventional inspection methods are often time-consuming, labor-dependent, and inefficient for large-scale photovoltaic plants. This study presents a computer vision framework based on deep learning for detecting and localizing surface regions in solar panel images. The proposed model integrates a ResNet101 feature extractor, a Sobel-based edge extraction module, a spatial–channel attention mechanism, an FPN/PAN multi-scale feature fusion structure, and a multi-scale detection head. The output consists of bounding boxes, region classes, and confidence scores, enabling simultaneous detection of multiple regions within a single image. The dataset was constructed by combining publicly available images with manually collected samples and includes four classes: breakage, clean region, dust, and shadow. Experimental results show that the proposed model achieves a Precision of 0.954, Recall of 0.939, F1-score of 0.946, mAP@0.5 of 0.966, and mAP@0.5:0.95 of 0.766, outperforming baseline photovoltaic defect detection models. The ablation study further demonstrates that edge information, spatial–channel attention, and multi-scale feature fusion substantially improve detection and localization accuracy. Overall, the proposed method provides a practical solution for automatic monitoring, targeted cleaning, inspection prioritization, and predictive maintenance of solar panels.