Semantic Division Network for Vehicle Image Classification

The patent changes the architectural parameters of the deep learning network by replacing standard convolutional layers with depth-wise separable convolutional layers. This parameter change reduces computational complexity and power consumption while maintaining recognition performance through the separable convolution structure that decomposes the convolution operation into depth-wise convolution and pointwise convolution steps.

The patent segments the convolutional neural network into distinct encoder and decoder units with specific layers. The encoder unit includes multiple convolutional layers followed by pooling layers, while the decoder unit uses transposed convolutional layers. This segmentation allows optimized processing where each unit performs specific functions, reducing overall computational burden while maintaining performance.

The patent changes the computational parameters by using depth-wise separable convolution instead of standard convolution. This parameter change makes the network more adaptable to simple hardware accelerators like MMA by reducing the computational intensity and memory bandwidth requirements, while still achieving high recognition performance that was previously only attainable on GPUs.

The patent employs a simplified network architecture that can run on cheaper, less powerful hardware accelerators rather than requiring expensive GPUs. The depth-wise separable convolution structure enables the system to achieve acceptable performance on lower-cost hardware, making the solution more versatile and cost-effective for embedded automotive applications.

The patent changes the convolution operation parameters by implementing depth-wise separable convolution, which divides the standard convolution into two separate operations: depth-wise convolution that processes each input channel independently, and pointwise convolution that combines the results. This parameter change reduces the number of multiplications and increases processing throughput while maintaining feature extraction quality.

The patent segments the convolutional processing into distinct stages within each layer: depth-wise convolution stage followed by pointwise convolution stage. This segmentation allows for more efficient computation where the depth-wise convolution captures spatial patterns and the pointwise convolution combines channel information, improving overall processing throughput without sacrificing feature quality.

The patent changes the convolutional layer structure by using depth-wise separable convolution, which reduces the number of parameters and computations required per layer. This parameter change allows the network to achieve good feature representation with fewer effective layers, reducing overall network complexity and memory requirements while maintaining recognition performance.

The patent discards the need for many heavy convolutional layers by using the more efficient depth-wise separable convolution structure. Instead of adding more standard convolutional layers to improve feature representation, the system recovers performance through the optimized convolution structure and strategic placement of pooling and upsampling layers, achieving the same representation quality with reduced complexity.