Academia.eduAcademia.edu

Figure 9 – uploaded by Dũng Võ

See full PDFdownloadDownload figure
Fig. 9. Luminance PSNR versus bitrate using H.264.

Figure 9 Luminance PSNR versus bitrate using H.264.

Related Figures (14)

Fig. 1. Dichromatic model. [3], the authors take a unique approach based on the human-vi- sual-system model by loosely using a Retinex method [18], [19] while operating in the discrete cosine transform (DCT) domain. Most of the transform-domain techniques may be applied either at the encoding end, i.e., Pre, or the decoding end, i.e., Post, but the performance comparisons are not explored.
Fig. 1. Dichromatic model. [3], the authors take a unique approach based on the human-vi- sual-system model by loosely using a Retinex method [18], [19] while operating in the discrete cosine transform (DCT) domain. Most of the transform-domain techniques may be applied either at the encoding end, i.e., Pre, or the decoding end, i.e., Post, but the performance comparisons are not explored.
As shown in Fig. 2(a), the closer pixels (less atmospheric veiling [14]) are darker than hazier background pixels that are farther away. Notice that, comparing (15) with (16), even if there was one foreground pixel at the extreme edge of neighborhood Q(m, n), the DCP will choose the foreground pixel value & ¢ at (m,n), whereas the MDCP method will choose the background pixel Zy. If no refinement is used, then this will cause halo arti- facts at occlusion boundaries for the case of using the DCP. In
As shown in Fig. 2(a), the closer pixels (less atmospheric veiling [14]) are darker than hazier background pixels that are farther away. Notice that, comparing (15) with (16), even if there was one foreground pixel at the extreme edge of neighborhood Q(m, n), the DCP will choose the foreground pixel value & ¢ at (m,n), whereas the MDCP method will choose the background pixel Zy. If no refinement is used, then this will cause halo arti- facts at occlusion boundaries for the case of using the DCP. In
Fig. 2. Examples of occlusion boundaries. (a) Original image. (b) Dehazed using unrefined DCP method (9). (c) Dehazed using our proposed MDCP method (12). (d) Dehazed using method from [14]. Notice in (b) and (d) that there are halo artifacts around the leaf edges. The halo artifacts are not as dominant in (c). (a) Hazy image x. (b) xp. (c) xy. (d) xr.
Fig. 2. Examples of occlusion boundaries. (a) Original image. (b) Dehazed using unrefined DCP method (9). (c) Dehazed using our proposed MDCP method (12). (d) Dehazed using method from [14]. Notice in (b) and (d) that there are halo artifacts around the leaf edges. The halo artifacts are not as dominant in (c). (a) Hazy image x. (b) xp. (c) xy. (d) xr.
The next step is to apply an N x N forward type-2 DCT. In the direct form, the (7,7)th N x N block of the transformed luminance is In [12], the dichromatic model is used to explain the contrast at isodepths (smooth depth regions) due to haze or fog being reduced. If we measure the gradient of the isodepth, threshold the result, and sum them, i.e., ||V - ||, we get a contrast measure. Therefore, we have the following relationship: Once the luminance is obtained, the values are offset to be within the range of [-2",2” — 1] for (P + 1) bit data. For clarity purposes, we will not make a notational difference and, from this point on, assume that the luminance values for 7% and y are subtracted by 2”. This brings us to the junction between the “Level and Offset” and “ N x N Block DCT” blocks in Fig. 3.
The next step is to apply an N x N forward type-2 DCT. In the direct form, the (7,7)th N x N block of the transformed luminance is In [12], the dichromatic model is used to explain the contrast at isodepths (smooth depth regions) due to haze or fog being reduced. If we measure the gradient of the isodepth, threshold the result, and sum them, i.e., ||V - ||, we get a contrast measure. Therefore, we have the following relationship: Once the luminance is obtained, the values are offset to be within the range of [-2",2” — 1] for (P + 1) bit data. For clarity purposes, we will not make a notational difference and, from this point on, assume that the luminance values for 7% and y are subtracted by 2”. This brings us to the junction between the “Level and Offset” and “ N x N Block DCT” blocks in Fig. 3.
Fig. 4. Comparison of the dehazed image using Pre and Post methods with similar file sizes. (a) Synthesized Haze. (b) Dehazed (no compression). (Red box) Region of interest used for visual comparison. (c) Uncompressed dehazed image. (d) Pre method. (e) Post method. image for the Post method. A cropped region of the results is dis- played in Fig. 4. Although we set the quality low enough such that the Pre method resulted in a smaller image size, the Post method suffered more from ringing artifacts as evident around the building edges in Fig. 4(e). We can also see more blocking artifacts in the sky region above the buildings (see Fig. 5).
Fig. 4. Comparison of the dehazed image using Pre and Post methods with similar file sizes. (a) Synthesized Haze. (b) Dehazed (no compression). (Red box) Region of interest used for visual comparison. (c) Uncompressed dehazed image. (d) Pre method. (e) Post method. image for the Post method. A cropped region of the results is dis- played in Fig. 4. Although we set the quality low enough such that the Pre method resulted in a smaller image size, the Post method suffered more from ringing artifacts as evident around the building edges in Fig. 4(e). We can also see more blocking artifacts in the sky region above the buildings (see Fig. 5).
Fig. 7. Performance comparison of the deblocked Pre and Post methods using JPEG and WILD. Fig. 6. Measure of blocking artifacts for each method at different image sizes.
Fig. 7. Performance comparison of the deblocked Pre and Post methods using JPEG and WILD. Fig. 6. Measure of blocking artifacts for each method at different image sizes.
Fig. 5. Real-world example of Pre and Post using MDCP (12) and JPEG compression. (a) Original image. (b) Post method with MDCP. (c) Pre method with MDCP. (d) Detected block artifacts in Post image using [26]. (e) Detected block artifacts in Pre image using [26].
Fig. 5. Real-world example of Pre and Post using MDCP (12) and JPEG compression. (a) Original image. (b) Post method with MDCP. (c) Pre method with MDCP. (d) Detected block artifacts in Post image using [26]. (e) Detected block artifacts in Pre image using [26].
“OMPARISON OF AVERAGE MV. AND RESIDUAL BITS PER FRAME Table I. The bit ratio for each method, i.e., Post/Pre, is added to the table to make it easy to see the bit allocation relationships.
“OMPARISON OF AVERAGE MV. AND RESIDUAL BITS PER FRAME Table I. The bit ratio for each method, i.e., Post/Pre, is added to the table to make it easy to see the bit allocation relationships.
Fig. 8. (a) Frame from the original video sequence. (b) Snapshot from the de- hazed video. (Red box) Region used for closer comparisons. [(c)—(e)] Compar- isons of the Pre and Post methods with H.264 compression at 250- and 500-kb/s bitrates. (a) Original video frame. (b) Dehazed video frame. (Red box) Region of interest used for comparisons. (c) 250-kb/s Post. (d) 250-kb/s Pre. (e) 500-kb/s Post. (f) 500-kb/s Pre.
Fig. 8. (a) Frame from the original video sequence. (b) Snapshot from the de- hazed video. (Red box) Region used for closer comparisons. [(c)—(e)] Compar- isons of the Pre and Post methods with H.264 compression at 250- and 500-kb/s bitrates. (a) Original video frame. (b) Dehazed video frame. (Red box) Region of interest used for comparisons. (c) 250-kb/s Post. (d) 250-kb/s Pre. (e) 500-kb/s Post. (f) 500-kb/s Pre.
Fig. 10. Measure of blocking artifacts for each method and bitrate. The metric was applied to every frame in a video sequence and then averaged.
Fig. 10. Measure of blocking artifacts for each method and bitrate. The metric was applied to every frame in a video sequence and then averaged.
over the Post method for any bitrate. The error-bar plot (using the rankings from Table I) of the ITU-R results in Fig. 11 shows two significant overlaps between the uncompressed (ranked first) and 1000-kb/s Pre sequences (ranked second). and the 250-kb/s Pre (ranked fifth) and 1000-kb/s Post (ranked sixth) sequences. The observers had difficulty distinguishing between the uncompressed enhanced sequence and the pre-en- hanced and compressed sequence. In addition, the observers had difficulty distinguishing between the 250-kb/s Pre and 1000-kb/s Post sequences. This is a significant compression difference by a factor of 4. If one is only concerned with a dehazed video sequence and can decide whether to enhance before or after compression, our results show choosing the Pre method can operate just as well as the traditionally chosen Posi method while operating in significantly lower bitrates.
over the Post method for any bitrate. The error-bar plot (using the rankings from Table I) of the ITU-R results in Fig. 11 shows two significant overlaps between the uncompressed (ranked first) and 1000-kb/s Pre sequences (ranked second). and the 250-kb/s Pre (ranked fifth) and 1000-kb/s Post (ranked sixth) sequences. The observers had difficulty distinguishing between the uncompressed enhanced sequence and the pre-en- hanced and compressed sequence. In addition, the observers had difficulty distinguishing between the 250-kb/s Pre and 1000-kb/s Post sequences. This is a significant compression difference by a factor of 4. If one is only concerned with a dehazed video sequence and can decide whether to enhance before or after compression, our results show choosing the Pre method can operate just as well as the traditionally chosen Posi method while operating in significantly lower bitrates.
Fig. 11. Plot of ITU-R-averaged marks with error bars versus rank. bits. The preserved details using Pre method contributed to the noticeable improvement in visual quality. We performed a subjective test with 20 different observers. These observers were presented with a test that displayed two videos side by side with all combinations such that the pool of randomly selected videos were different with respect to their bitrates and method (Pre or Post). After each playback, the ob- server was asked to compare the left video to the right. The choices the observers were allowed to select for comparing the left video sequence to the right were adjectival categorical judg- ments ranging from Much worse, Worse, Slightly worse, The same, Slightly better, Better, and Much better. This type of test is called the stimulus comparison adjectival categorical judgment (SCACJ) [36]. After processing the data, only the results from one observer was rejected due to rejection criteria in [36]. The remaining 19 observer results were used to compute the ITU-R statistics. For the pool of video sequences, we choose a range of bitrates that were close enough together to give us a good res- olution in our results but limited the maximum bitrate to 1000 kb/s because it was difficult beyond that to distinguish from the uncompressed sequence. — ArmlanmntT ay +1 414. rmmo4t14 #é6oTT ore es azed Video ITU-R Averaged Mark ITU-R Left Confidence Border ITU-R Right Confidence Border ITU-R Std. Dev
Fig. 11. Plot of ITU-R-averaged marks with error bars versus rank. bits. The preserved details using Pre method contributed to the noticeable improvement in visual quality. We performed a subjective test with 20 different observers. These observers were presented with a test that displayed two videos side by side with all combinations such that the pool of randomly selected videos were different with respect to their bitrates and method (Pre or Post). After each playback, the ob- server was asked to compare the left video to the right. The choices the observers were allowed to select for comparing the left video sequence to the right were adjectival categorical judg- ments ranging from Much worse, Worse, Slightly worse, The same, Slightly better, Better, and Much better. This type of test is called the stimulus comparison adjectival categorical judgment (SCACJ) [36]. After processing the data, only the results from one observer was rejected due to rejection criteria in [36]. The remaining 19 observer results were used to compute the ITU-R statistics. For the pool of video sequences, we choose a range of bitrates that were close enough together to give us a good res- olution in our results but limited the maximum bitrate to 1000 kb/s because it was difficult beyond that to distinguish from the uncompressed sequence. — ArmlanmntT ay +1 414. rmmo4t14 #é6oTT ore es azed Video ITU-R Averaged Mark ITU-R Left Confidence Border ITU-R Right Confidence Border ITU-R Std. Dev
efficient implementation.
efficient implementation.
Related topics:
Cognitive ScienceVideo CodingMotion estimationContrast-Enhanced UltrasoundStreaming MediaElectrical and Electronic EngineeringImage Color Analysis
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved