Hi Rokesh,
I understand that you want to know which convolution fitler size to choose for you deep learning architecture given the input image size is 566x804x3.
Choosing suitable convolution filter sizes can enhance the model's ability to capture relevant features effectively. Smaller filters like 3x3 or 5x5 are often ideal for capturing fine-grained details and edges, making them effective across many convolutional layers, especially when preserving spatial resolution through padding. In initial layers, medium-sized filters like 7x7 or 11x11 can be useful to capture more complex shapes and textures over a larger receptive field, providing a good balance between detail and context. Larger filters, such as 13x13 or 15x15, may be applied in the very first layer if the image is high-resolution; these filters help capture broad, context-rich features but increase computational load and reduce spatial dimensions faster.
Rather than applying a single filter size throughout, it is beneficial to start with larger filters and progressively reduce to smaller ones in deeper layers, as seen in architectures like VGG.
Experimenting with a mix of filter sizes can improve feature extraction efficiency and model performance, especially in high-resolution or complex images.
Hope this helps.
