Integer-Quantized ResNet In-Loop Filters for Lower Video Coding Complexity

The patent applies parameter changes by transforming the transformer block operators from floating-point arithmetic to integer arithmetic. This is achieved through rounding floating-point values to nearest integers and using integer quantization techniques. The parameter transformation maintains the functional behavior of the transformer operators while enabling efficient integer-based computation, thereby reducing computational complexity without significant degradation of coding quality.

The patent substitutes the mechanical computation system by replacing floating-point arithmetic operations with integer arithmetic operations in the transformer blocks. This substitution is implemented through integer quantization of the transformer operators, which allows the system to perform the same filtering function using simpler integer math instead of complex floating-point math, thus reducing computational burden while maintaining acceptable coding performance.

The patent carefully designs the integer quantization parameters to achieve an optimal balance between computational complexity and coding quality. By adjusting the quantization precision and rounding strategies, the patent reduces computational complexity through integer arithmetic while minimizing quality degradation. The quantization parameters are optimized to retain sufficient information for effective in-loop filtering.

The patent applies partial quantization selectively to specific transformer operators rather than all operations, or uses asymmetric quantization where different precision levels are applied to different parts of the computation. This partial action approach reduces overall computational complexity while maintaining coding quality in critical regions where full precision is most beneficial.