Feature Information Fusion and Lightweight ResNet for Image Semantic Segmentation

We introduce an efficient segmentation framework that integrates multi-scale cues in real time, avoiding the need for overly deep architectures. A Separable Pyramid Module (SPM) is introduced to harvest rich context at 1/4 and 1/8 resolutions by combining depthwise-separable, factorized and dilated convolutions in a bottleneck layout, cutting parameters while preserving receptive fields. To guide the fusion of high-level semantics into low-level detail, a Context Channel Attention (CCA) block is proposed. It re-weights shallow feature channels by exploiting the inter-channel correlations learned from deep feature maps, refining edges without extra heavy computation. The overall encoder-decoder is deliberately kept shallow, so that the deepest feature map remains at 1/8 scale, ensuring fast inference. Extensive experiments on PASCAL VOC2012 demonstrate that the new method achieves competitive accuracy against deeper counterparts while maintaining superior speed, validating the effectiveness of the SPM and CCA designs for balancing precision and real-time performance in semantic segmentation tasks.