**6. Results and discussion**

Considering the work in the literature [60] and in order to test the proposed codesign, we used the same accelerated CNNs (CNN\_A, CNN\_B) with different layers and configuration summarized in **Table 1**. Th parameters are summarized in different data precision for **w**eights and **f**eature maps. We test then the proposed QoS-QoR aware CNN co-design on an object detection. The suggested co-design schemes finds the most promising CNN topology example for the intended hardware system and application as a bundle containing depth-wise Cnv3 (DW-Cnv3), point-wise Conv1 (PW-Cnv3), and max-pooling layers. Depending on this data, the co-design investigates 3 CNN configurations, each with a distinct normalization strategy, in order to meet the QoR and QoS requirements. **Table 1** shows the different result of the proposed scheme. Using the FPGA Pynq Z1 and the proposed architecture achieved a best results when used different CNNs (CNN\_A, CNN\_B). According to these results, CNN\_A occupies 27% FFs, 78% BRAMs, 84% DSPs, and 76% LUTs with a working frequency of 150 MHz. In addition it reached a 23 FPS with a maximum latency of 44 ms, a maximum power of about 2.6 W and an energy efficiency of about 0.114 J/ image. On the other hand, the CNN\_B with the configuration of W16 & F8 occupies a 38% FFs, 96% BRAMs, 91% DSPs, and 83% LUTs with a working frequency of 150 MHz. According to this highest hardware cost compared to the first topology, CNN\_B cannot surpass the first one considering its energy efficiency factor which is of about 0.16 J/image.
