<p>The high-dimensional characteristics inherent in the light field confer significant advantages for tasks related to three-dimensional scene-aware understanding. However, any degradation in the quality of angular or spatial dimensions can profoundly impact visual analysis tasks. Quality assessment methods for light field images lack an analysis of the intrinsic cross-influence of distortions in both the angular and spatial domains. This paper designs convolutional layers customized for light field cross characteristics in angular and spatial domains, enabling multi-level feature extraction and fusion. We construct feature extraction layers from macro-pixel patches, ensuring that each training patch contains both angular and spatial information. A convolutional neural network-based angular-spatial feature extraction model is proposed, which has angular and spatial feature extraction layers, angular-spatial feature fusion layers, and dual-task layers. Angular feature extraction realizes the sparsification of the angular dimension of input patches and obtains preliminary angular-spatial feature fusion. The spatial features are extracted from sparse angular cells using dilated convolution. Deep angular-spatial fusion is conducted to learn the cross-influence of distortion in the angular and spatial domains. Followed with a dual-task training module for distortion levels and quality scores. The proposed model has an order of magnitude fewer parameters than state-of-the-art algorithms, while simultaneously achieving excellent performance, as verified by extensive comparative experiments across up to six subjective databases.</p>