Existing CNN streaming accelerators largely overlook inter-layer pipeline imbalance, which reduces resource efficiency by causing idle processing time. To address this, we propose a low-cost, high-throughput framework that transforms CNN models into pipeline-balanced architectures. It uses the Roofline model for initial bottleneck analysis and a layer-wise resource allocation algorithm to dynamically balance computational loads, optimizing throughput and resource utilization. We validate our proposed framework on the Intel Arria 10 GX1150 FPGA, demonstrating that multiple CNN models achieve throughput and energy efficiency comparable to state-of-the-art CNN accelerators. Compared to a baseline CNN accelerator, our approach achieves up to 5.1× performance improvement on the same network, while DSP resource utilization efficiency improves by up to 1.74× over previous works. Furthermore, based on extensive empirical analysis, we provide design guidelines for evaluating the suitability of CNN models for streaming architectures, offering practical insights to assist researchers in optimizing CNN models at the algorithmic level or selecting the most appropriate models for deployment in streaming architectures.